mirror_zfs/module/zfs/spa_log_spacemap.c

1323 lines
47 KiB
C
Raw Normal View History

Log Spacemap Project = Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent db587941c5ff6dea01932bb78f70db63cf7f38ba Update vdev_is_spacemap_addressable() for new spacemap encoding 419ba5914552c6185afbe1dd17b3ed4b0d526547 Simplify spa_sync by breaking it up to smaller functions 8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834 Factor metaslab_load_wait() in metaslab_load() b194fab0fb6caad18711abccaff3c69ad8b3f6d3 Rename range_tree_verify to range_tree_verify_not_present df72b8bebe0ebac0b20e0750984bad182cb6564a Change target size of metaslabs from 256GB to 16GB c853f382db731e15a87512f4ef1101d14d778a55 zdb -L should skip leak detection altogether 21e7cf5da89f55ce98ec1115726b150e19eefe89 vs_alloc can underflow in L2ARC vdevs 7558997d2f808368867ca7e5234e5793446e8f3f Simplify log vdev removal code 6c926f426a26ffb6d7d8e563e33fc176164175cb Get rid of space_map_update() for ms_synced_length 425d3237ee88abc53d8522a7139c926d278b4b7f Introduce auxiliary metaslab histograms 928e8ad47d3478a3d5d01f0dd6ae74a9371af65e Error path in metaslab_load_impl() forgets to drop ms_sync_lock 8eef997679ba54547f7d361553d21b3291f41ae7 = References Background, Motivation, and Internals of the Feature - OpenZFS 2017 Presentation: youtu.be/jj2IxRkl5bQ - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project Flushing Algorithm Internals & Performance Results (Illumos Specific) - Blogpost: sdimitro.github.io/post/zfs-lsm-flushing/ - OpenZFS 2018 Presentation: youtu.be/x6D2dHRjkxw - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm Upstream Delphix Issues: DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320 DLPX-63385 Reviewed-by: Sean Eric Fagan <sef@ixsystems.com> Reviewed-by: Matt Ahrens <matt@delphix.com> Reviewed-by: George Wilson <gwilson@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com> Closes #8442
2019-07-16 20:11:49 +03:00
/*
* CDDL HEADER START
*
* The contents of this file are subject to the terms of the
* Common Development and Distribution License (the "License").
* You may not use this file except in compliance with the License.
*
* You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
* or http://www.opensolaris.org/os/licensing.
* See the License for the specific language governing permissions
* and limitations under the License.
*
* When distributing Covered Code, include this CDDL HEADER in each
* file and include the License file at usr/src/OPENSOLARIS.LICENSE.
* If applicable, add the following below this CDDL HEADER, with the
* fields enclosed by brackets "[]" replaced with your own identifying
* information: Portions Copyright [yyyy] [name of copyright owner]
*
* CDDL HEADER END
*/
/*
* Copyright (c) 2018, 2019 by Delphix. All rights reserved.
*/
#include <sys/dmu_objset.h>
#include <sys/metaslab.h>
#include <sys/metaslab_impl.h>
#include <sys/spa.h>
#include <sys/spa_impl.h>
#include <sys/spa_log_spacemap.h>
#include <sys/vdev_impl.h>
#include <sys/zap.h>
/*
* Log Space Maps
*
* Log space maps are an optimization in ZFS metadata allocations for pools
* whose workloads are primarily random-writes. Random-write workloads are also
* typically random-free, meaning that they are freeing from locations scattered
* throughout the pool. This means that each TXG we will have to append some
* FREE records to almost every metaslab. With log space maps, we hold their
* changes in memory and log them altogether in one pool-wide space map on-disk
* for persistence. As more blocks are accumulated in the log space maps and
* more unflushed changes are accounted in memory, we flush a selected group
* of metaslabs every TXG to relieve memory pressure and potential overheads
* when loading the pool. Flushing a metaslab to disk relieves memory as we
* flush any unflushed changes from memory to disk (i.e. the metaslab's space
* map) and saves import time by making old log space maps obsolete and
* eventually destroying them. [A log space map is said to be obsolete when all
* its entries have made it to their corresponding metaslab space maps].
*
* == On disk data structures used ==
*
* - The pool has a new feature flag and a new entry in the MOS. The feature
* is activated when we create the first log space map and remains active
* for the lifetime of the pool. The new entry in the MOS Directory [refer
* to DMU_POOL_LOG_SPACEMAP_ZAP] is populated with a ZAP whose key-value
* pairs are of the form <key: txg, value: log space map object for that txg>.
* This entry is our on-disk reference of the log space maps that exist in
* the pool for each TXG and it is used during import to load all the
* metaslab unflushed changes in memory. To see how this structure is first
* created and later populated refer to spa_generate_syncing_log_sm(). To see
* how it is used during import time refer to spa_ld_log_sm_metadata().
*
* - Each vdev has a new entry in its vdev_top_zap (see field
* VDEV_TOP_ZAP_MS_UNFLUSHED_PHYS_TXGS) which holds the msp_unflushed_txg of
* each metaslab in this vdev. This field is the on-disk counterpart of the
* in-memory field ms_unflushed_txg which tells us from which TXG and onwards
* the metaslab haven't had its changes flushed. During import, we use this
* to ignore any entries in the space map log that are for this metaslab but
* from a TXG before msp_unflushed_txg. At that point, we also populate its
* in-memory counterpart and from there both fields are updated every time
* we flush that metaslab.
*
* - A space map is created every TXG and, during that TXG, it is used to log
* all incoming changes (the log space map). When created, the log space map
* is referenced in memory by spa_syncing_log_sm and its object ID is inserted
* to the space map ZAP mentioned above. The log space map is closed at the
* end of the TXG and will be destroyed when it becomes fully obsolete. We
* know when a log space map has become obsolete by looking at the oldest
* (and smallest) ms_unflushed_txg in the pool. If the value of that is bigger
* than the log space map's TXG, then it means that there is no metaslab who
* doesn't have the changes from that log and we can therefore destroy it.
* [see spa_cleanup_old_sm_logs()].
*
* == Important in-memory structures ==
*
* - The per-spa field spa_metaslabs_by_flushed sorts all the metaslabs in
* the pool by their ms_unflushed_txg field. It is primarily used for three
* reasons. First of all, it is used during flushing where we try to flush
* metaslabs in-order from the oldest-flushed to the most recently flushed
* every TXG. Secondly, it helps us to lookup the ms_unflushed_txg of the
* oldest flushed metaslab to distinguish which log space maps have become
* obsolete and which ones are still relevant. Finally it tells us which
* metaslabs have unflushed changes in a pool where this feature was just
* enabled, as we don't immediately add all of the pool's metaslabs but we
* add them over time as they go through metaslab_sync(). The reason that
* we do that is to ease these pools into the behavior of the flushing
* algorithm (described later on).
*
* - The per-spa field spa_sm_logs_by_txg can be thought as the in-memory
* counterpart of the space map ZAP mentioned above. It's an AVL tree whose
* nodes represent the log space maps in the pool. This in-memory
* representation of log space maps in the pool sorts the log space maps by
* the TXG that they were created (which is also the TXG of their unflushed
* changes). It also contains the following extra information for each
* space map:
* [1] The number of metaslabs that were last flushed on that TXG. This is
* important because if that counter is zero and this is the oldest
* log then it means that it is also obsolete.
* [2] The number of blocks of that space map. This field is used by the
* block heuristic of our flushing algorithm (described later on).
* It represents how many blocks of metadata changes ZFS had to write
* to disk for that TXG.
*
* - The per-spa field spa_log_summary is a list of entries that summarizes
* the metaslab and block counts of all the nodes of the spa_sm_logs_by_txg
* AVL tree mentioned above. The reason this exists is that our flushing
* algorithm (described later) tries to estimate how many metaslabs to flush
* in each TXG by iterating over all the log space maps and looking at their
* block counts. Summarizing that information means that don't have to
* iterate through each space map, minimizing the runtime overhead of the
* flushing algorithm which would be induced in syncing context. In terms of
* implementation the log summary is used as a queue:
* * we modify or pop entries from its head when we flush metaslabs
* * we modify or append entries to its tail when we sync changes.
*
* - Each metaslab has two new range trees that hold its unflushed changes,
* ms_unflushed_allocs and ms_unflushed_frees. These are always disjoint.
*
* == Flushing algorithm ==
*
* The decision of how many metaslabs to flush on a give TXG is guided by
* two heuristics:
*
* [1] The memory heuristic -
* We keep track of the memory used by the unflushed trees from all the
* metaslabs [see sus_memused of spa_unflushed_stats] and we ensure that it
* stays below a certain threshold which is determined by an arbitrary hard
* limit and an arbitrary percentage of the system's memory [see
* spa_log_exceeds_memlimit()]. When we see that the memory usage of the
* unflushed changes are passing that threshold, we flush metaslabs, which
* empties their unflushed range trees, reducing the memory used.
*
* [2] The block heuristic -
* We try to keep the total number of blocks in the log space maps in check
* so the log doesn't grow indefinitely and we don't induce a lot of overhead
* when loading the pool. At the same time we don't want to flush a lot of
* metaslabs too often as this would defeat the purpose of the log space map.
* As a result we set a limit in the amount of blocks that we think it's
* acceptable for the log space maps to have and try not to cross it.
* [see sus_blocklimit from spa_unflushed_stats].
*
* In order to stay below the block limit every TXG we have to estimate how
* many metaslabs we need to flush based on the current rate of incoming blocks
* and our history of log space map blocks. The main idea here is to answer
* the question of how many metaslabs do we need to flush in order to get rid
* at least an X amount of log space map blocks. We can answer this question
* by iterating backwards from the oldest log space map to the newest one
* and looking at their metaslab and block counts. At this point the log summary
* mentioned above comes handy as it reduces the amount of things that we have
* to iterate (even though it may reduce the preciseness of our estimates due
* to its aggregation of data). So with that in mind, we project the incoming
* rate of the current TXG into the future and attempt to approximate how many
* metaslabs would we need to flush from now in order to avoid exceeding our
* block limit in different points in the future (granted that we would keep
* flushing the same number of metaslabs for every TXG). Then we take the
* maximum number from all these estimates to be on the safe side. For the
* exact implementation details of algorithm refer to
* spa_estimate_metaslabs_to_flush.
*/
/*
* This is used as the block size for the space maps used for the
* log space map feature. These space maps benefit from a bigger
* block size as we expect to be writing a lot of data to them at
* once.
*/
unsigned long zfs_log_sm_blksz = 1ULL << 17;
/*
* Percentage of the overall system's memory that ZFS allows to be
Log Spacemap Project = Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent db587941c5ff6dea01932bb78f70db63cf7f38ba Update vdev_is_spacemap_addressable() for new spacemap encoding 419ba5914552c6185afbe1dd17b3ed4b0d526547 Simplify spa_sync by breaking it up to smaller functions 8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834 Factor metaslab_load_wait() in metaslab_load() b194fab0fb6caad18711abccaff3c69ad8b3f6d3 Rename range_tree_verify to range_tree_verify_not_present df72b8bebe0ebac0b20e0750984bad182cb6564a Change target size of metaslabs from 256GB to 16GB c853f382db731e15a87512f4ef1101d14d778a55 zdb -L should skip leak detection altogether 21e7cf5da89f55ce98ec1115726b150e19eefe89 vs_alloc can underflow in L2ARC vdevs 7558997d2f808368867ca7e5234e5793446e8f3f Simplify log vdev removal code 6c926f426a26ffb6d7d8e563e33fc176164175cb Get rid of space_map_update() for ms_synced_length 425d3237ee88abc53d8522a7139c926d278b4b7f Introduce auxiliary metaslab histograms 928e8ad47d3478a3d5d01f0dd6ae74a9371af65e Error path in metaslab_load_impl() forgets to drop ms_sync_lock 8eef997679ba54547f7d361553d21b3291f41ae7 = References Background, Motivation, and Internals of the Feature - OpenZFS 2017 Presentation: youtu.be/jj2IxRkl5bQ - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project Flushing Algorithm Internals & Performance Results (Illumos Specific) - Blogpost: sdimitro.github.io/post/zfs-lsm-flushing/ - OpenZFS 2018 Presentation: youtu.be/x6D2dHRjkxw - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm Upstream Delphix Issues: DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320 DLPX-63385 Reviewed-by: Sean Eric Fagan <sef@ixsystems.com> Reviewed-by: Matt Ahrens <matt@delphix.com> Reviewed-by: George Wilson <gwilson@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com> Closes #8442
2019-07-16 20:11:49 +03:00
* used for unflushed changes (e.g. the sum of size of all the nodes
* in the unflushed trees).
*
* Note that this value is calculated over 1000000 for finer granularity
* (thus the _ppm suffix; reads as "parts per million"). As an example,
* the default of 1000 allows 0.1% of memory to be used.
*/
unsigned long zfs_unflushed_max_mem_ppm = 1000;
/*
* Specific hard-limit in memory that ZFS allows to be used for
* unflushed changes.
*/
unsigned long zfs_unflushed_max_mem_amt = 1ULL << 30;
/*
* The following tunable determines the number of blocks that can be used for
* the log space maps. It is expressed as a percentage of the total number of
* metaslabs in the pool (i.e. the default of 400 means that the number of log
* blocks is capped at 4 times the number of metaslabs).
*
* This value exists to tune our flushing algorithm, with higher values
* flushing metaslabs less often (doing less I/Os) per TXG versus lower values
* flushing metaslabs more aggressively with the upside of saving overheads
* when loading the pool. Another factor in this tradeoff is that flushing
* less often can potentially lead to better utilization of the metaslab space
* map's block size as we accumulate more changes per flush.
*
* Given that this tunable indirectly controls the flush rate (metaslabs
* flushed per txg) and that's why making it a percentage in terms of the
* number of metaslabs in the pool makes sense here.
*
* As a rule of thumb we default this tunable to 400% based on the following:
*
* 1] Assuming a constant flush rate and a constant incoming rate of log blocks
* it is reasonable to expect that the amount of obsolete entries changes
* linearly from txg to txg (e.g. the oldest log should have the most
* obsolete entries, and the most recent one the least). With this we could
* say that, at any given time, about half of the entries in the whole space
* map log are obsolete. Thus for every two entries for a metaslab in the
* log space map, only one of them is valid and actually makes it to the
* metaslab's space map.
* [factor of 2]
* 2] Each entry in the log space map is guaranteed to be two words while
* entries in metaslab space maps are generally single-word.
* [an extra factor of 2 - 400% overall]
* 3] Even if [1] and [2] are slightly less than 2 each, we haven't taken into
* account any consolidation of segments from the log space map to the
* unflushed range trees nor their history (e.g. a segment being allocated,
* then freed, then allocated again means 3 log space map entries but 0
* metaslab space map entries). Depending on the workload, we've seen ~1.8
* non-obsolete log space map entries per metaslab entry, for a total of
* ~600%. Since most of these estimates though are workload dependent, we
* default on 400% to be conservative.
*
* Thus we could say that even in the worst
* case of [1] and [2], the factor should end up being 4.
*
* That said, regardless of the number of metaslabs in the pool we need to
* provide upper and lower bounds for the log block limit.
* [see zfs_unflushed_log_block_{min,max}]
*/
unsigned long zfs_unflushed_log_block_pct = 400;
/*
* If the number of metaslabs is small and our incoming rate is high, we could
* get into a situation that we are flushing all our metaslabs every TXG. Thus
* we always allow at least this many log blocks.
*/
unsigned long zfs_unflushed_log_block_min = 1000;
/*
* If the log becomes too big, the import time of the pool can take a hit in
* terms of performance. Thus we have a hard limit in the size of the log in
* terms of blocks.
*/
unsigned long zfs_unflushed_log_block_max = (1ULL << 18);
/*
* Max # of rows allowed for the log_summary. The tradeoff here is accuracy and
* stability of the flushing algorithm (longer summary) vs its runtime overhead
* (smaller summary is faster to traverse).
*/
unsigned long zfs_max_logsm_summary_length = 10;
/*
* Tunable that sets the lower bound on the metaslabs to flush every TXG.
*
* Setting this to 0 has no effect since if the pool is idle we won't even be
* creating log space maps and therefore we won't be flushing. On the other
* hand if the pool has any incoming workload our block heuristic will start
* flushing metaslabs anyway.
*
* The point of this tunable is to be used in extreme cases where we really
* want to flush more metaslabs than our adaptable heuristic plans to flush.
*/
unsigned long zfs_min_metaslabs_to_flush = 1;
/*
* Tunable that specifies how far in the past do we want to look when trying to
* estimate the incoming log blocks for the current TXG.
*
* Setting this too high may not only increase runtime but also minimize the
* effect of the incoming rates from the most recent TXGs as we take the
* average over all the blocks that we walk
* [see spa_estimate_incoming_log_blocks].
*/
unsigned long zfs_max_log_walking = 5;
/*
* This tunable exists solely for testing purposes. It ensures that the log
* spacemaps are not flushed and destroyed during export in order for the
* relevant log spacemap import code paths to be tested (effectively simulating
* a crash).
*/
int zfs_keep_log_spacemaps_at_export = 0;
static uint64_t
spa_estimate_incoming_log_blocks(spa_t *spa)
{
ASSERT3U(spa_sync_pass(spa), ==, 1);
uint64_t steps = 0, sum = 0;
for (spa_log_sm_t *sls = avl_last(&spa->spa_sm_logs_by_txg);
sls != NULL && steps < zfs_max_log_walking;
sls = AVL_PREV(&spa->spa_sm_logs_by_txg, sls)) {
if (sls->sls_txg == spa_syncing_txg(spa)) {
/*
* skip the log created in this TXG as this would
* make our estimations inaccurate.
*/
continue;
}
sum += sls->sls_nblocks;
steps++;
}
return ((steps > 0) ? DIV_ROUND_UP(sum, steps) : 0);
}
uint64_t
spa_log_sm_blocklimit(spa_t *spa)
{
return (spa->spa_unflushed_stats.sus_blocklimit);
}
void
spa_log_sm_set_blocklimit(spa_t *spa)
{
if (!spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP)) {
ASSERT0(spa_log_sm_blocklimit(spa));
return;
}
uint64_t calculated_limit =
(spa_total_metaslabs(spa) * zfs_unflushed_log_block_pct) / 100;
spa->spa_unflushed_stats.sus_blocklimit = MIN(MAX(calculated_limit,
zfs_unflushed_log_block_min), zfs_unflushed_log_block_max);
}
uint64_t
spa_log_sm_nblocks(spa_t *spa)
{
return (spa->spa_unflushed_stats.sus_nblocks);
}
/*
* Ensure that the in-memory log space map structures and the summary
* have the same block and metaslab counts.
*/
static void
spa_log_summary_verify_counts(spa_t *spa)
{
ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP));
if ((zfs_flags & ZFS_DEBUG_LOG_SPACEMAP) == 0)
return;
uint64_t ms_in_avl = avl_numnodes(&spa->spa_metaslabs_by_flushed);
uint64_t ms_in_summary = 0, blk_in_summary = 0;
for (log_summary_entry_t *e = list_head(&spa->spa_log_summary);
e; e = list_next(&spa->spa_log_summary, e)) {
ms_in_summary += e->lse_mscount;
blk_in_summary += e->lse_blkcount;
}
uint64_t ms_in_logs = 0, blk_in_logs = 0;
for (spa_log_sm_t *sls = avl_first(&spa->spa_sm_logs_by_txg);
sls; sls = AVL_NEXT(&spa->spa_sm_logs_by_txg, sls)) {
ms_in_logs += sls->sls_mscount;
blk_in_logs += sls->sls_nblocks;
}
VERIFY3U(ms_in_logs, ==, ms_in_summary);
VERIFY3U(ms_in_logs, ==, ms_in_avl);
VERIFY3U(blk_in_logs, ==, blk_in_summary);
VERIFY3U(blk_in_logs, ==, spa_log_sm_nblocks(spa));
}
static boolean_t
summary_entry_is_full(spa_t *spa, log_summary_entry_t *e)
{
uint64_t blocks_per_row = MAX(1,
DIV_ROUND_UP(spa_log_sm_blocklimit(spa),
zfs_max_logsm_summary_length));
return (blocks_per_row <= e->lse_blkcount);
}
/*
* Update the log summary information to reflect the fact that a metaslab
* was flushed or destroyed (e.g due to device removal or pool export/destroy).
*
* We typically flush the oldest flushed metaslab so the first (and oldest)
Log Spacemap Project = Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent db587941c5ff6dea01932bb78f70db63cf7f38ba Update vdev_is_spacemap_addressable() for new spacemap encoding 419ba5914552c6185afbe1dd17b3ed4b0d526547 Simplify spa_sync by breaking it up to smaller functions 8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834 Factor metaslab_load_wait() in metaslab_load() b194fab0fb6caad18711abccaff3c69ad8b3f6d3 Rename range_tree_verify to range_tree_verify_not_present df72b8bebe0ebac0b20e0750984bad182cb6564a Change target size of metaslabs from 256GB to 16GB c853f382db731e15a87512f4ef1101d14d778a55 zdb -L should skip leak detection altogether 21e7cf5da89f55ce98ec1115726b150e19eefe89 vs_alloc can underflow in L2ARC vdevs 7558997d2f808368867ca7e5234e5793446e8f3f Simplify log vdev removal code 6c926f426a26ffb6d7d8e563e33fc176164175cb Get rid of space_map_update() for ms_synced_length 425d3237ee88abc53d8522a7139c926d278b4b7f Introduce auxiliary metaslab histograms 928e8ad47d3478a3d5d01f0dd6ae74a9371af65e Error path in metaslab_load_impl() forgets to drop ms_sync_lock 8eef997679ba54547f7d361553d21b3291f41ae7 = References Background, Motivation, and Internals of the Feature - OpenZFS 2017 Presentation: youtu.be/jj2IxRkl5bQ - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project Flushing Algorithm Internals & Performance Results (Illumos Specific) - Blogpost: sdimitro.github.io/post/zfs-lsm-flushing/ - OpenZFS 2018 Presentation: youtu.be/x6D2dHRjkxw - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm Upstream Delphix Issues: DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320 DLPX-63385 Reviewed-by: Sean Eric Fagan <sef@ixsystems.com> Reviewed-by: Matt Ahrens <matt@delphix.com> Reviewed-by: George Wilson <gwilson@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com> Closes #8442
2019-07-16 20:11:49 +03:00
* entry of the summary is updated. However if that metaslab is getting loaded
* we may flush the second oldest one which may be part of an entry later in
* the summary. Moreover, if we call into this function from metaslab_fini()
* the metaslabs probably won't be ordered by ms_unflushed_txg. Thus we ask
* for a txg as an argument so we can locate the appropriate summary entry for
* the metaslab.
*/
void
spa_log_summary_decrement_mscount(spa_t *spa, uint64_t txg)
{
/*
* We don't track summary data for read-only pools and this function
* can be called from metaslab_fini(). In that case return immediately.
*/
if (!spa_writeable(spa))
return;
log_summary_entry_t *target = NULL;
for (log_summary_entry_t *e = list_head(&spa->spa_log_summary);
e != NULL; e = list_next(&spa->spa_log_summary, e)) {
if (e->lse_start > txg)
break;
target = e;
}
if (target == NULL || target->lse_mscount == 0) {
/*
* We didn't find a summary entry for this metaslab. We must be
* at the teardown of a spa_load() attempt that got an error
* while reading the log space maps.
*/
VERIFY3S(spa_load_state(spa), ==, SPA_LOAD_ERROR);
return;
}
target->lse_mscount--;
}
/*
* Update the log summary information to reflect the fact that we destroyed
* old log space maps. Since we can only destroy the oldest log space maps,
* we decrement the block count of the oldest summary entry and potentially
* destroy it when that count hits 0.
*
* This function is called after a metaslab is flushed and typically that
* metaslab is the oldest flushed, which means that this function will
* typically decrement the block count of the first entry of the summary and
* potentially free it if the block count gets to zero (its metaslab count
* should be zero too at that point).
*
* There are certain scenarios though that don't work exactly like that so we
* need to account for them:
*
* Scenario [1]: It is possible that after we flushed the oldest flushed
* metaslab and we destroyed the oldest log space map, more recent logs had 0
* metaslabs pointing to them so we got rid of them too. This can happen due
* to metaslabs being destroyed through device removal, or because the oldest
* flushed metaslab was loading but we kept flushing more recently flushed
* metaslabs due to the memory pressure of unflushed changes. Because of that,
* we always iterate from the beginning of the summary and if blocks_gone is
* bigger than the block_count of the current entry we free that entry (we
* expect its metaslab count to be zero), we decrement blocks_gone and on to
* the next entry repeating this procedure until blocks_gone gets decremented
* to 0. Doing this also works for the typical case mentioned above.
*
* Scenario [2]: The oldest flushed metaslab isn't necessarily accounted by
* the first (and oldest) entry in the summary. If the first few entries of
* the summary were only accounting metaslabs from a device that was just
* removed, then the current oldest flushed metaslab could be accounted by an
* entry somewhere in the middle of the summary. Moreover flushing that
* metaslab will destroy all the log space maps older than its ms_unflushed_txg
* because they became obsolete after the removal. Thus, iterating as we did
* for scenario [1] works out for this case too.
*
* Scenario [3]: At times we decide to flush all the metaslabs in the pool
* in one TXG (either because we are exporting the pool or because our flushing
* heuristics decided to do so). When that happens all the log space maps get
* destroyed except the one created for the current TXG which doesn't have
* any log blocks yet. As log space maps get destroyed with every metaslab that
* we flush, entries in the summary are also destroyed. This brings a weird
* corner-case when we flush the last metaslab and the log space map of the
* current TXG is in the same summary entry with other log space maps that
* are older. When that happens we are eventually left with this one last
* summary entry whose blocks are gone (blocks_gone equals the entry's block
* count) but its metaslab count is non-zero (because it accounts all the
* metaslabs in the pool as they all got flushed). Under this scenario we can't
* free this last summary entry as it's referencing all the metaslabs in the
* pool and its block count will get incremented at the end of this sync (when
* we close the syncing log space map). Thus we just decrement its current
* block count and leave it alone. In the case that the pool gets exported,
* its metaslab count will be decremented over time as we call metaslab_fini()
* for all the metaslabs in the pool and the entry will be freed at
* spa_unload_log_sm_metadata().
*/
void
spa_log_summary_decrement_blkcount(spa_t *spa, uint64_t blocks_gone)
{
for (log_summary_entry_t *e = list_head(&spa->spa_log_summary);
e != NULL; e = list_head(&spa->spa_log_summary)) {
if (e->lse_blkcount > blocks_gone) {
/*
* Assert that we stopped at an entry that is not
* obsolete.
*/
ASSERT(e->lse_mscount != 0);
e->lse_blkcount -= blocks_gone;
blocks_gone = 0;
break;
} else if (e->lse_mscount == 0) {
/* remove obsolete entry */
blocks_gone -= e->lse_blkcount;
list_remove(&spa->spa_log_summary, e);
kmem_free(e, sizeof (log_summary_entry_t));
} else {
/* Verify that this is scenario [3] mentioned above. */
VERIFY3U(blocks_gone, ==, e->lse_blkcount);
/*
* Assert that this is scenario [3] further by ensuring
* that this is the only entry in the summary.
*/
VERIFY3P(e, ==, list_tail(&spa->spa_log_summary));
ASSERT3P(e, ==, list_head(&spa->spa_log_summary));
blocks_gone = e->lse_blkcount = 0;
break;
}
}
/*
* Ensure that there is no way we are trying to remove more blocks
* than the # of blocks in the summary.
*/
ASSERT0(blocks_gone);
}
void
spa_log_sm_decrement_mscount(spa_t *spa, uint64_t txg)
{
spa_log_sm_t target = { .sls_txg = txg };
spa_log_sm_t *sls = avl_find(&spa->spa_sm_logs_by_txg,
&target, NULL);
if (sls == NULL) {
/*
* We must be at the teardown of a spa_load() attempt that
* got an error while reading the log space maps.
*/
VERIFY3S(spa_load_state(spa), ==, SPA_LOAD_ERROR);
return;
}
ASSERT(sls->sls_mscount > 0);
sls->sls_mscount--;
}
void
spa_log_sm_increment_current_mscount(spa_t *spa)
{
spa_log_sm_t *last_sls = avl_last(&spa->spa_sm_logs_by_txg);
ASSERT3U(last_sls->sls_txg, ==, spa_syncing_txg(spa));
last_sls->sls_mscount++;
}
static void
summary_add_data(spa_t *spa, uint64_t txg, uint64_t metaslabs_flushed,
uint64_t nblocks)
{
log_summary_entry_t *e = list_tail(&spa->spa_log_summary);
if (e == NULL || summary_entry_is_full(spa, e)) {
e = kmem_zalloc(sizeof (log_summary_entry_t), KM_SLEEP);
e->lse_start = txg;
list_insert_tail(&spa->spa_log_summary, e);
}
ASSERT3U(e->lse_start, <=, txg);
e->lse_mscount += metaslabs_flushed;
e->lse_blkcount += nblocks;
}
static void
spa_log_summary_add_incoming_blocks(spa_t *spa, uint64_t nblocks)
{
summary_add_data(spa, spa_syncing_txg(spa), 0, nblocks);
}
void
spa_log_summary_add_flushed_metaslab(spa_t *spa)
{
summary_add_data(spa, spa_syncing_txg(spa), 1, 0);
}
/*
* This function attempts to estimate how many metaslabs should
* we flush to satisfy our block heuristic for the log spacemap
* for the upcoming TXGs.
*
* Specifically, it first tries to estimate the number of incoming
* blocks in this TXG. Then by projecting that incoming rate to
* future TXGs and using the log summary, it figures out how many
* flushes we would need to do for future TXGs individually to
* stay below our block limit and returns the maximum number of
* flushes from those estimates.
*/
static uint64_t
spa_estimate_metaslabs_to_flush(spa_t *spa)
{
ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP));
ASSERT3U(spa_sync_pass(spa), ==, 1);
ASSERT(spa_log_sm_blocklimit(spa) != 0);
/*
* This variable contains the incoming rate that will be projected
* and used for our flushing estimates in the future.
*/
uint64_t incoming = spa_estimate_incoming_log_blocks(spa);
/*
* At any point in time this variable tells us how many
* TXGs in the future we are so we can make our estimations.
*/
uint64_t txgs_in_future = 1;
/*
* This variable tells us how much room do we have until we hit
* our limit. When it goes negative, it means that we've exceeded
* our limit and we need to flush.
*
* Note that since we start at the first TXG in the future (i.e.
* txgs_in_future starts from 1) we already decrement this
* variable by the incoming rate.
*/
int64_t available_blocks =
spa_log_sm_blocklimit(spa) - spa_log_sm_nblocks(spa) - incoming;
/*
* This variable tells us the total number of flushes needed to
* keep the log size within the limit when we reach txgs_in_future.
*/
uint64_t total_flushes = 0;
/* Holds the current maximum of our estimates so far. */
uint64_t max_flushes_pertxg =
MIN(avl_numnodes(&spa->spa_metaslabs_by_flushed),
zfs_min_metaslabs_to_flush);
/*
* For our estimations we only look as far in the future
* as the summary allows us.
*/
for (log_summary_entry_t *e = list_head(&spa->spa_log_summary);
e; e = list_next(&spa->spa_log_summary, e)) {
/*
* If there is still room before we exceed our limit
* then keep skipping TXGs accumulating more blocks
* based on the incoming rate until we exceed it.
*/
if (available_blocks >= 0) {
uint64_t skip_txgs = (available_blocks / incoming) + 1;
available_blocks -= (skip_txgs * incoming);
txgs_in_future += skip_txgs;
ASSERT3S(available_blocks, >=, -incoming);
}
/*
* At this point we're far enough into the future where
* the limit was just exceeded and we flush metaslabs
* based on the current entry in the summary, updating
* our available_blocks.
*/
ASSERT3S(available_blocks, <, 0);
available_blocks += e->lse_blkcount;
total_flushes += e->lse_mscount;
/*
* Keep the running maximum of the total_flushes that
* we've done so far over the number of TXGs in the
* future that we are. The idea here is to estimate
* the average number of flushes that we should do
* every TXG so that when we are that many TXGs in the
* future we stay under the limit.
*/
max_flushes_pertxg = MAX(max_flushes_pertxg,
DIV_ROUND_UP(total_flushes, txgs_in_future));
ASSERT3U(avl_numnodes(&spa->spa_metaslabs_by_flushed), >=,
max_flushes_pertxg);
}
return (max_flushes_pertxg);
}
uint64_t
spa_log_sm_memused(spa_t *spa)
{
return (spa->spa_unflushed_stats.sus_memused);
}
static boolean_t
spa_log_exceeds_memlimit(spa_t *spa)
{
if (spa_log_sm_memused(spa) > zfs_unflushed_max_mem_amt)
return (B_TRUE);
uint64_t system_mem_allowed = ((physmem * PAGESIZE) *
zfs_unflushed_max_mem_ppm) / 1000000;
if (spa_log_sm_memused(spa) > system_mem_allowed)
return (B_TRUE);
return (B_FALSE);
}
boolean_t
spa_flush_all_logs_requested(spa_t *spa)
{
return (spa->spa_log_flushall_txg != 0);
}
void
spa_flush_metaslabs(spa_t *spa, dmu_tx_t *tx)
{
uint64_t txg = dmu_tx_get_txg(tx);
if (spa_sync_pass(spa) != 1)
return;
if (!spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP))
return;
/*
* If we don't have any metaslabs with unflushed changes
* return immediately.
*/
if (avl_numnodes(&spa->spa_metaslabs_by_flushed) == 0)
return;
/*
* During SPA export we leave a few empty TXGs to go by [see
* spa_final_dirty_txg() to understand why]. For this specific
* case, it is important to not flush any metaslabs as that
* would dirty this TXG.
*
* That said, during one of these dirty TXGs that is less or
* equal to spa_final_dirty(), spa_unload() will request that
* we try to flush all the metaslabs for that TXG before
* exporting the pool, thus we ensure that we didn't get a
* request of flushing everything before we attempt to return
* immediately.
*/
if (spa->spa_uberblock.ub_rootbp.blk_birth < txg &&
!dmu_objset_is_dirty(spa_meta_objset(spa), txg) &&
!spa_flush_all_logs_requested(spa))
return;
/*
* We need to generate a log space map before flushing because this
* will set up the in-memory data (i.e. node in spa_sm_logs_by_txg)
* for this TXG's flushed metaslab count (aka sls_mscount which is
* manipulated in many ways down the metaslab_flush() codepath).
*
* That is not to say that we may generate a log space map when we
* don't need it. If we are flushing metaslabs, that means that we
* were going to write changes to disk anyway, so even if we were
* not flushing, a log space map would have been created anyway in
* metaslab_sync().
*/
spa_generate_syncing_log_sm(spa, tx);
/*
* This variable tells us how many metaslabs we want to flush based
* on the block-heuristic of our flushing algorithm (see block comment
* of log space map feature). We also decrement this as we flush
* metaslabs and attempt to destroy old log space maps.
*/
uint64_t want_to_flush;
if (spa_flush_all_logs_requested(spa)) {
ASSERT3S(spa_state(spa), ==, POOL_STATE_EXPORTED);
want_to_flush = avl_numnodes(&spa->spa_metaslabs_by_flushed);
} else {
want_to_flush = spa_estimate_metaslabs_to_flush(spa);
}
ASSERT3U(avl_numnodes(&spa->spa_metaslabs_by_flushed), >=,
want_to_flush);
/* Used purely for verification purposes */
uint64_t visited = 0;
/*
* Ideally we would only iterate through spa_metaslabs_by_flushed
* using only one variable (curr). We can't do that because
* metaslab_flush() mutates position of curr in the AVL when
* it flushes that metaslab by moving it to the end of the tree.
* Thus we always keep track of the original next node of the
* current node (curr) in another variable (next).
*/
metaslab_t *next = NULL;
for (metaslab_t *curr = avl_first(&spa->spa_metaslabs_by_flushed);
curr != NULL; curr = next) {
next = AVL_NEXT(&spa->spa_metaslabs_by_flushed, curr);
/*
* If this metaslab has been flushed this txg then we've done
* a full circle over the metaslabs.
*/
if (metaslab_unflushed_txg(curr) == txg)
break;
/*
* If we are done flushing for the block heuristic and the
* unflushed changes don't exceed the memory limit just stop.
*/
if (want_to_flush == 0 && !spa_log_exceeds_memlimit(spa))
break;
mutex_enter(&curr->ms_sync_lock);
mutex_enter(&curr->ms_lock);
boolean_t flushed = metaslab_flush(curr, tx);
mutex_exit(&curr->ms_lock);
mutex_exit(&curr->ms_sync_lock);
/*
* If we failed to flush a metaslab (because it was loading),
* then we are done with the block heuristic as it's not
* possible to destroy any log space maps once you've skipped
* a metaslab. In that case we just set our counter to 0 but
* we continue looping in case there is still memory pressure
* due to unflushed changes. Note that, flushing a metaslab
* that is not the oldest flushed in the pool, will never
* destroy any log space maps [see spa_cleanup_old_sm_logs()].
*/
if (!flushed) {
want_to_flush = 0;
} else if (want_to_flush > 0) {
want_to_flush--;
}
visited++;
}
ASSERT3U(avl_numnodes(&spa->spa_metaslabs_by_flushed), >=, visited);
}
/*
* Close the log space map for this TXG and update the block counts
* for the log's in-memory structure and the summary.
Log Spacemap Project = Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent db587941c5ff6dea01932bb78f70db63cf7f38ba Update vdev_is_spacemap_addressable() for new spacemap encoding 419ba5914552c6185afbe1dd17b3ed4b0d526547 Simplify spa_sync by breaking it up to smaller functions 8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834 Factor metaslab_load_wait() in metaslab_load() b194fab0fb6caad18711abccaff3c69ad8b3f6d3 Rename range_tree_verify to range_tree_verify_not_present df72b8bebe0ebac0b20e0750984bad182cb6564a Change target size of metaslabs from 256GB to 16GB c853f382db731e15a87512f4ef1101d14d778a55 zdb -L should skip leak detection altogether 21e7cf5da89f55ce98ec1115726b150e19eefe89 vs_alloc can underflow in L2ARC vdevs 7558997d2f808368867ca7e5234e5793446e8f3f Simplify log vdev removal code 6c926f426a26ffb6d7d8e563e33fc176164175cb Get rid of space_map_update() for ms_synced_length 425d3237ee88abc53d8522a7139c926d278b4b7f Introduce auxiliary metaslab histograms 928e8ad47d3478a3d5d01f0dd6ae74a9371af65e Error path in metaslab_load_impl() forgets to drop ms_sync_lock 8eef997679ba54547f7d361553d21b3291f41ae7 = References Background, Motivation, and Internals of the Feature - OpenZFS 2017 Presentation: youtu.be/jj2IxRkl5bQ - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project Flushing Algorithm Internals & Performance Results (Illumos Specific) - Blogpost: sdimitro.github.io/post/zfs-lsm-flushing/ - OpenZFS 2018 Presentation: youtu.be/x6D2dHRjkxw - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm Upstream Delphix Issues: DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320 DLPX-63385 Reviewed-by: Sean Eric Fagan <sef@ixsystems.com> Reviewed-by: Matt Ahrens <matt@delphix.com> Reviewed-by: George Wilson <gwilson@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com> Closes #8442
2019-07-16 20:11:49 +03:00
*/
void
spa_sync_close_syncing_log_sm(spa_t *spa)
{
if (spa_syncing_log_sm(spa) == NULL)
return;
ASSERT(spa_feature_is_active(spa, SPA_FEATURE_LOG_SPACEMAP));
spa_log_sm_t *sls = avl_last(&spa->spa_sm_logs_by_txg);
ASSERT3U(sls->sls_txg, ==, spa_syncing_txg(spa));
sls->sls_nblocks = space_map_nblocks(spa_syncing_log_sm(spa));
spa->spa_unflushed_stats.sus_nblocks += sls->sls_nblocks;
/*
* Note that we can't assert that sls_mscount is not 0,
* because there is the case where the first metaslab
* in spa_metaslabs_by_flushed is loading and we were
* not able to flush any metaslabs the current TXG.
*/
ASSERT(sls->sls_nblocks != 0);
spa_log_summary_add_incoming_blocks(spa, sls->sls_nblocks);
spa_log_summary_verify_counts(spa);
space_map_close(spa->spa_syncing_log_sm);
spa->spa_syncing_log_sm = NULL;
/*
* At this point we tried to flush as many metaslabs as we
* can as the pool is getting exported. Reset the "flush all"
* so the last few TXGs before closing the pool can be empty
* (e.g. not dirty).
*/
if (spa_flush_all_logs_requested(spa)) {
ASSERT3S(spa_state(spa), ==, POOL_STATE_EXPORTED);
spa->spa_log_flushall_txg = 0;
}
}
void
spa_cleanup_old_sm_logs(spa_t *spa, dmu_tx_t *tx)
{
objset_t *mos = spa_meta_objset(spa);
uint64_t spacemap_zap;
int error = zap_lookup(mos, DMU_POOL_DIRECTORY_OBJECT,
DMU_POOL_LOG_SPACEMAP_ZAP, sizeof (spacemap_zap), 1, &spacemap_zap);
if (error == ENOENT) {
ASSERT(avl_is_empty(&spa->spa_sm_logs_by_txg));
return;
}
VERIFY0(error);
metaslab_t *oldest = avl_first(&spa->spa_metaslabs_by_flushed);
uint64_t oldest_flushed_txg = metaslab_unflushed_txg(oldest);
/* Free all log space maps older than the oldest_flushed_txg. */
for (spa_log_sm_t *sls = avl_first(&spa->spa_sm_logs_by_txg);
sls && sls->sls_txg < oldest_flushed_txg;
sls = avl_first(&spa->spa_sm_logs_by_txg)) {
ASSERT0(sls->sls_mscount);
avl_remove(&spa->spa_sm_logs_by_txg, sls);
space_map_free_obj(mos, sls->sls_sm_obj, tx);
VERIFY0(zap_remove_int(mos, spacemap_zap, sls->sls_txg, tx));
spa->spa_unflushed_stats.sus_nblocks -= sls->sls_nblocks;
kmem_free(sls, sizeof (spa_log_sm_t));
}
}
static spa_log_sm_t *
spa_log_sm_alloc(uint64_t sm_obj, uint64_t txg)
{
spa_log_sm_t *sls = kmem_zalloc(sizeof (*sls), KM_SLEEP);
sls->sls_sm_obj = sm_obj;
sls->sls_txg = txg;
return (sls);
}
void
spa_generate_syncing_log_sm(spa_t *spa, dmu_tx_t *tx)
{
uint64_t txg = dmu_tx_get_txg(tx);
objset_t *mos = spa_meta_objset(spa);
if (spa_syncing_log_sm(spa) != NULL)
return;
if (!spa_feature_is_enabled(spa, SPA_FEATURE_LOG_SPACEMAP))
return;
uint64_t spacemap_zap;
int error = zap_lookup(mos, DMU_POOL_DIRECTORY_OBJECT,
DMU_POOL_LOG_SPACEMAP_ZAP, sizeof (spacemap_zap), 1, &spacemap_zap);
if (error == ENOENT) {
ASSERT(avl_is_empty(&spa->spa_sm_logs_by_txg));
error = 0;
spacemap_zap = zap_create(mos,
DMU_OTN_ZAP_METADATA, DMU_OT_NONE, 0, tx);
VERIFY0(zap_add(mos, DMU_POOL_DIRECTORY_OBJECT,
DMU_POOL_LOG_SPACEMAP_ZAP, sizeof (spacemap_zap), 1,
&spacemap_zap, tx));
spa_feature_incr(spa, SPA_FEATURE_LOG_SPACEMAP, tx);
}
VERIFY0(error);
uint64_t sm_obj;
ASSERT3U(zap_lookup_int_key(mos, spacemap_zap, txg, &sm_obj),
==, ENOENT);
sm_obj = space_map_alloc(mos, zfs_log_sm_blksz, tx);
VERIFY0(zap_add_int_key(mos, spacemap_zap, txg, sm_obj, tx));
avl_add(&spa->spa_sm_logs_by_txg, spa_log_sm_alloc(sm_obj, txg));
/*
* We pass UINT64_MAX as the space map's representation size
* and SPA_MINBLOCKSHIFT as the shift, to make the space map
* accept any sorts of segments since there's no real advantage
* to being more restrictive (given that we're already going
* to be using 2-word entries).
*/
VERIFY0(space_map_open(&spa->spa_syncing_log_sm, mos, sm_obj,
0, UINT64_MAX, SPA_MINBLOCKSHIFT));
/*
* If the log space map feature was just enabled, the blocklimit
* has not yet been set.
*/
if (spa_log_sm_blocklimit(spa) == 0)
spa_log_sm_set_blocklimit(spa);
}
/*
* Find all the log space maps stored in the space map ZAP and sort
* them by their TXG in spa_sm_logs_by_txg.
*/
static int
spa_ld_log_sm_metadata(spa_t *spa)
{
int error;
uint64_t spacemap_zap;
ASSERT(avl_is_empty(&spa->spa_sm_logs_by_txg));
error = zap_lookup(spa_meta_objset(spa), DMU_POOL_DIRECTORY_OBJECT,
DMU_POOL_LOG_SPACEMAP_ZAP, sizeof (spacemap_zap), 1, &spacemap_zap);
if (error == ENOENT) {
/* the space map ZAP doesn't exist yet */
return (0);
} else if (error != 0) {
spa_load_failed(spa, "spa_ld_log_sm_metadata(): failed at "
Log Spacemap Project = Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent db587941c5ff6dea01932bb78f70db63cf7f38ba Update vdev_is_spacemap_addressable() for new spacemap encoding 419ba5914552c6185afbe1dd17b3ed4b0d526547 Simplify spa_sync by breaking it up to smaller functions 8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834 Factor metaslab_load_wait() in metaslab_load() b194fab0fb6caad18711abccaff3c69ad8b3f6d3 Rename range_tree_verify to range_tree_verify_not_present df72b8bebe0ebac0b20e0750984bad182cb6564a Change target size of metaslabs from 256GB to 16GB c853f382db731e15a87512f4ef1101d14d778a55 zdb -L should skip leak detection altogether 21e7cf5da89f55ce98ec1115726b150e19eefe89 vs_alloc can underflow in L2ARC vdevs 7558997d2f808368867ca7e5234e5793446e8f3f Simplify log vdev removal code 6c926f426a26ffb6d7d8e563e33fc176164175cb Get rid of space_map_update() for ms_synced_length 425d3237ee88abc53d8522a7139c926d278b4b7f Introduce auxiliary metaslab histograms 928e8ad47d3478a3d5d01f0dd6ae74a9371af65e Error path in metaslab_load_impl() forgets to drop ms_sync_lock 8eef997679ba54547f7d361553d21b3291f41ae7 = References Background, Motivation, and Internals of the Feature - OpenZFS 2017 Presentation: youtu.be/jj2IxRkl5bQ - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project Flushing Algorithm Internals & Performance Results (Illumos Specific) - Blogpost: sdimitro.github.io/post/zfs-lsm-flushing/ - OpenZFS 2018 Presentation: youtu.be/x6D2dHRjkxw - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm Upstream Delphix Issues: DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320 DLPX-63385 Reviewed-by: Sean Eric Fagan <sef@ixsystems.com> Reviewed-by: Matt Ahrens <matt@delphix.com> Reviewed-by: George Wilson <gwilson@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com> Closes #8442
2019-07-16 20:11:49 +03:00
"zap_lookup(DMU_POOL_DIRECTORY_OBJECT) [error %d]",
error);
return (error);
}
zap_cursor_t zc;
zap_attribute_t za;
for (zap_cursor_init(&zc, spa_meta_objset(spa), spacemap_zap);
(error = zap_cursor_retrieve(&zc, &za)) == 0;
zap_cursor_advance(&zc)) {
Log Spacemap Project = Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent db587941c5ff6dea01932bb78f70db63cf7f38ba Update vdev_is_spacemap_addressable() for new spacemap encoding 419ba5914552c6185afbe1dd17b3ed4b0d526547 Simplify spa_sync by breaking it up to smaller functions 8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834 Factor metaslab_load_wait() in metaslab_load() b194fab0fb6caad18711abccaff3c69ad8b3f6d3 Rename range_tree_verify to range_tree_verify_not_present df72b8bebe0ebac0b20e0750984bad182cb6564a Change target size of metaslabs from 256GB to 16GB c853f382db731e15a87512f4ef1101d14d778a55 zdb -L should skip leak detection altogether 21e7cf5da89f55ce98ec1115726b150e19eefe89 vs_alloc can underflow in L2ARC vdevs 7558997d2f808368867ca7e5234e5793446e8f3f Simplify log vdev removal code 6c926f426a26ffb6d7d8e563e33fc176164175cb Get rid of space_map_update() for ms_synced_length 425d3237ee88abc53d8522a7139c926d278b4b7f Introduce auxiliary metaslab histograms 928e8ad47d3478a3d5d01f0dd6ae74a9371af65e Error path in metaslab_load_impl() forgets to drop ms_sync_lock 8eef997679ba54547f7d361553d21b3291f41ae7 = References Background, Motivation, and Internals of the Feature - OpenZFS 2017 Presentation: youtu.be/jj2IxRkl5bQ - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project Flushing Algorithm Internals & Performance Results (Illumos Specific) - Blogpost: sdimitro.github.io/post/zfs-lsm-flushing/ - OpenZFS 2018 Presentation: youtu.be/x6D2dHRjkxw - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm Upstream Delphix Issues: DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320 DLPX-63385 Reviewed-by: Sean Eric Fagan <sef@ixsystems.com> Reviewed-by: Matt Ahrens <matt@delphix.com> Reviewed-by: George Wilson <gwilson@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com> Closes #8442
2019-07-16 20:11:49 +03:00
uint64_t log_txg = zfs_strtonum(za.za_name, NULL);
spa_log_sm_t *sls =
spa_log_sm_alloc(za.za_first_integer, log_txg);
avl_add(&spa->spa_sm_logs_by_txg, sls);
}
zap_cursor_fini(&zc);
if (error != ENOENT) {
spa_load_failed(spa, "spa_ld_log_sm_metadata(): failed at "
"zap_cursor_retrieve(spacemap_zap) [error %d]",
error);
return (error);
}
Log Spacemap Project = Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent db587941c5ff6dea01932bb78f70db63cf7f38ba Update vdev_is_spacemap_addressable() for new spacemap encoding 419ba5914552c6185afbe1dd17b3ed4b0d526547 Simplify spa_sync by breaking it up to smaller functions 8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834 Factor metaslab_load_wait() in metaslab_load() b194fab0fb6caad18711abccaff3c69ad8b3f6d3 Rename range_tree_verify to range_tree_verify_not_present df72b8bebe0ebac0b20e0750984bad182cb6564a Change target size of metaslabs from 256GB to 16GB c853f382db731e15a87512f4ef1101d14d778a55 zdb -L should skip leak detection altogether 21e7cf5da89f55ce98ec1115726b150e19eefe89 vs_alloc can underflow in L2ARC vdevs 7558997d2f808368867ca7e5234e5793446e8f3f Simplify log vdev removal code 6c926f426a26ffb6d7d8e563e33fc176164175cb Get rid of space_map_update() for ms_synced_length 425d3237ee88abc53d8522a7139c926d278b4b7f Introduce auxiliary metaslab histograms 928e8ad47d3478a3d5d01f0dd6ae74a9371af65e Error path in metaslab_load_impl() forgets to drop ms_sync_lock 8eef997679ba54547f7d361553d21b3291f41ae7 = References Background, Motivation, and Internals of the Feature - OpenZFS 2017 Presentation: youtu.be/jj2IxRkl5bQ - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project Flushing Algorithm Internals & Performance Results (Illumos Specific) - Blogpost: sdimitro.github.io/post/zfs-lsm-flushing/ - OpenZFS 2018 Presentation: youtu.be/x6D2dHRjkxw - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm Upstream Delphix Issues: DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320 DLPX-63385 Reviewed-by: Sean Eric Fagan <sef@ixsystems.com> Reviewed-by: Matt Ahrens <matt@delphix.com> Reviewed-by: George Wilson <gwilson@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com> Closes #8442
2019-07-16 20:11:49 +03:00
for (metaslab_t *m = avl_first(&spa->spa_metaslabs_by_flushed);
m; m = AVL_NEXT(&spa->spa_metaslabs_by_flushed, m)) {
spa_log_sm_t target = { .sls_txg = metaslab_unflushed_txg(m) };
spa_log_sm_t *sls = avl_find(&spa->spa_sm_logs_by_txg,
&target, NULL);
/*
* At this point if sls is zero it means that a bug occurred
* in ZFS the last time the pool was open or earlier in the
* import code path. In general, we would have placed a
* VERIFY() here or in this case just let the kernel panic
* with NULL pointer dereference when incrementing sls_mscount,
* but since this is the import code path we can be a bit more
* lenient. Thus, for DEBUG bits we always cause a panic, while
* in production we log the error and just fail the import.
*/
ASSERT(sls != NULL);
if (sls == NULL) {
spa_load_failed(spa, "spa_ld_log_sm_metadata(): bug "
"encountered: could not find log spacemap for "
"TXG %ld [error %d]",
metaslab_unflushed_txg(m), ENOENT);
return (ENOENT);
}
Log Spacemap Project = Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent db587941c5ff6dea01932bb78f70db63cf7f38ba Update vdev_is_spacemap_addressable() for new spacemap encoding 419ba5914552c6185afbe1dd17b3ed4b0d526547 Simplify spa_sync by breaking it up to smaller functions 8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834 Factor metaslab_load_wait() in metaslab_load() b194fab0fb6caad18711abccaff3c69ad8b3f6d3 Rename range_tree_verify to range_tree_verify_not_present df72b8bebe0ebac0b20e0750984bad182cb6564a Change target size of metaslabs from 256GB to 16GB c853f382db731e15a87512f4ef1101d14d778a55 zdb -L should skip leak detection altogether 21e7cf5da89f55ce98ec1115726b150e19eefe89 vs_alloc can underflow in L2ARC vdevs 7558997d2f808368867ca7e5234e5793446e8f3f Simplify log vdev removal code 6c926f426a26ffb6d7d8e563e33fc176164175cb Get rid of space_map_update() for ms_synced_length 425d3237ee88abc53d8522a7139c926d278b4b7f Introduce auxiliary metaslab histograms 928e8ad47d3478a3d5d01f0dd6ae74a9371af65e Error path in metaslab_load_impl() forgets to drop ms_sync_lock 8eef997679ba54547f7d361553d21b3291f41ae7 = References Background, Motivation, and Internals of the Feature - OpenZFS 2017 Presentation: youtu.be/jj2IxRkl5bQ - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project Flushing Algorithm Internals & Performance Results (Illumos Specific) - Blogpost: sdimitro.github.io/post/zfs-lsm-flushing/ - OpenZFS 2018 Presentation: youtu.be/x6D2dHRjkxw - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm Upstream Delphix Issues: DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320 DLPX-63385 Reviewed-by: Sean Eric Fagan <sef@ixsystems.com> Reviewed-by: Matt Ahrens <matt@delphix.com> Reviewed-by: George Wilson <gwilson@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com> Closes #8442
2019-07-16 20:11:49 +03:00
sls->sls_mscount++;
}
return (0);
}
typedef struct spa_ld_log_sm_arg {
spa_t *slls_spa;
uint64_t slls_txg;
} spa_ld_log_sm_arg_t;
static int
spa_ld_log_sm_cb(space_map_entry_t *sme, void *arg)
{
uint64_t offset = sme->sme_offset;
uint64_t size = sme->sme_run;
uint32_t vdev_id = sme->sme_vdev;
spa_ld_log_sm_arg_t *slls = arg;
spa_t *spa = slls->slls_spa;
vdev_t *vd = vdev_lookup_top(spa, vdev_id);
/*
* If the vdev has been removed (i.e. it is indirect or a hole)
* skip this entry. The contents of this vdev have already moved
* elsewhere.
*/
if (!vdev_is_concrete(vd))
return (0);
metaslab_t *ms = vd->vdev_ms[offset >> vd->vdev_ms_shift];
ASSERT(!ms->ms_loaded);
/*
* If we have already flushed entries for this TXG to this
* metaslab's space map, then ignore it. Note that we flush
* before processing any allocations/frees for that TXG, so
* the metaslab's space map only has entries from *before*
* the unflushed TXG.
*/
if (slls->slls_txg < metaslab_unflushed_txg(ms))
return (0);
switch (sme->sme_type) {
case SM_ALLOC:
range_tree_remove_xor_add_segment(offset, offset + size,
ms->ms_unflushed_frees, ms->ms_unflushed_allocs);
break;
case SM_FREE:
range_tree_remove_xor_add_segment(offset, offset + size,
ms->ms_unflushed_allocs, ms->ms_unflushed_frees);
break;
default:
panic("invalid maptype_t");
break;
}
return (0);
}
static int
spa_ld_log_sm_data(spa_t *spa)
{
int error = 0;
/*
* If we are not going to do any writes there is no need
* to read the log space maps.
*/
if (!spa_writeable(spa))
return (0);
ASSERT0(spa->spa_unflushed_stats.sus_nblocks);
ASSERT0(spa->spa_unflushed_stats.sus_memused);
hrtime_t read_logs_starttime = gethrtime();
/* this is a no-op when we don't have space map logs */
for (spa_log_sm_t *sls = avl_first(&spa->spa_sm_logs_by_txg);
sls; sls = AVL_NEXT(&spa->spa_sm_logs_by_txg, sls)) {
space_map_t *sm = NULL;
error = space_map_open(&sm, spa_meta_objset(spa),
sls->sls_sm_obj, 0, UINT64_MAX, SPA_MINBLOCKSHIFT);
if (error != 0) {
spa_load_failed(spa, "spa_ld_log_sm_data(): failed at "
"space_map_open(obj=%llu) [error %d]",
(u_longlong_t)sls->sls_sm_obj, error);
goto out;
}
struct spa_ld_log_sm_arg vla = {
.slls_spa = spa,
.slls_txg = sls->sls_txg
};
error = space_map_iterate(sm, space_map_length(sm),
spa_ld_log_sm_cb, &vla);
if (error != 0) {
space_map_close(sm);
spa_load_failed(spa, "spa_ld_log_sm_data(): failed "
"at space_map_iterate(obj=%llu) [error %d]",
(u_longlong_t)sls->sls_sm_obj, error);
goto out;
}
ASSERT0(sls->sls_nblocks);
sls->sls_nblocks = space_map_nblocks(sm);
spa->spa_unflushed_stats.sus_nblocks += sls->sls_nblocks;
summary_add_data(spa, sls->sls_txg,
sls->sls_mscount, sls->sls_nblocks);
space_map_close(sm);
}
hrtime_t read_logs_endtime = gethrtime();
spa_load_note(spa,
"read %llu log space maps (%llu total blocks - blksz = %llu bytes) "
"in %lld ms", (u_longlong_t)avl_numnodes(&spa->spa_sm_logs_by_txg),
(u_longlong_t)spa_log_sm_nblocks(spa),
(u_longlong_t)zfs_log_sm_blksz,
(longlong_t)((read_logs_endtime - read_logs_starttime) / 1000000));
out:
/*
* Now that the metaslabs contain their unflushed changes:
* [1] recalculate their actual allocated space
* [2] recalculate their weights
* [3] sum up the memory usage of their unflushed range trees
* [4] optionally load them, if debug_load is set
*
* Note that even in the case where we get here because of an
* error (e.g. error != 0), we still want to update the fields
* below in order to have a proper teardown in spa_unload().
*/
for (metaslab_t *m = avl_first(&spa->spa_metaslabs_by_flushed);
m != NULL; m = AVL_NEXT(&spa->spa_metaslabs_by_flushed, m)) {
mutex_enter(&m->ms_lock);
m->ms_allocated_space = space_map_allocated(m->ms_sm) +
range_tree_space(m->ms_unflushed_allocs) -
range_tree_space(m->ms_unflushed_frees);
vdev_t *vd = m->ms_group->mg_vd;
metaslab_space_update(vd, m->ms_group->mg_class,
range_tree_space(m->ms_unflushed_allocs), 0, 0);
metaslab_space_update(vd, m->ms_group->mg_class,
-range_tree_space(m->ms_unflushed_frees), 0, 0);
ASSERT0(m->ms_weight & METASLAB_ACTIVE_MASK);
metaslab_recalculate_weight_and_sort(m);
spa->spa_unflushed_stats.sus_memused +=
metaslab_unflushed_changes_memused(m);
if (metaslab_debug_load && m->ms_sm != NULL) {
VERIFY0(metaslab_load(m));
metaslab_set_selected_txg(m, 0);
Log Spacemap Project = Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent db587941c5ff6dea01932bb78f70db63cf7f38ba Update vdev_is_spacemap_addressable() for new spacemap encoding 419ba5914552c6185afbe1dd17b3ed4b0d526547 Simplify spa_sync by breaking it up to smaller functions 8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834 Factor metaslab_load_wait() in metaslab_load() b194fab0fb6caad18711abccaff3c69ad8b3f6d3 Rename range_tree_verify to range_tree_verify_not_present df72b8bebe0ebac0b20e0750984bad182cb6564a Change target size of metaslabs from 256GB to 16GB c853f382db731e15a87512f4ef1101d14d778a55 zdb -L should skip leak detection altogether 21e7cf5da89f55ce98ec1115726b150e19eefe89 vs_alloc can underflow in L2ARC vdevs 7558997d2f808368867ca7e5234e5793446e8f3f Simplify log vdev removal code 6c926f426a26ffb6d7d8e563e33fc176164175cb Get rid of space_map_update() for ms_synced_length 425d3237ee88abc53d8522a7139c926d278b4b7f Introduce auxiliary metaslab histograms 928e8ad47d3478a3d5d01f0dd6ae74a9371af65e Error path in metaslab_load_impl() forgets to drop ms_sync_lock 8eef997679ba54547f7d361553d21b3291f41ae7 = References Background, Motivation, and Internals of the Feature - OpenZFS 2017 Presentation: youtu.be/jj2IxRkl5bQ - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project Flushing Algorithm Internals & Performance Results (Illumos Specific) - Blogpost: sdimitro.github.io/post/zfs-lsm-flushing/ - OpenZFS 2018 Presentation: youtu.be/x6D2dHRjkxw - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm Upstream Delphix Issues: DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320 DLPX-63385 Reviewed-by: Sean Eric Fagan <sef@ixsystems.com> Reviewed-by: Matt Ahrens <matt@delphix.com> Reviewed-by: George Wilson <gwilson@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com> Closes #8442
2019-07-16 20:11:49 +03:00
}
mutex_exit(&m->ms_lock);
}
return (error);
}
static int
spa_ld_unflushed_txgs(vdev_t *vd)
{
spa_t *spa = vd->vdev_spa;
objset_t *mos = spa_meta_objset(spa);
if (vd->vdev_top_zap == 0)
return (0);
uint64_t object = 0;
int error = zap_lookup(mos, vd->vdev_top_zap,
VDEV_TOP_ZAP_MS_UNFLUSHED_PHYS_TXGS,
sizeof (uint64_t), 1, &object);
if (error == ENOENT)
return (0);
else if (error != 0) {
spa_load_failed(spa, "spa_ld_unflushed_txgs(): failed at "
"zap_lookup(vdev_top_zap=%llu) [error %d]",
(u_longlong_t)vd->vdev_top_zap, error);
return (error);
}
for (uint64_t m = 0; m < vd->vdev_ms_count; m++) {
metaslab_t *ms = vd->vdev_ms[m];
ASSERT(ms != NULL);
metaslab_unflushed_phys_t entry;
uint64_t entry_size = sizeof (entry);
uint64_t entry_offset = ms->ms_id * entry_size;
error = dmu_read(mos, object,
entry_offset, entry_size, &entry, 0);
if (error != 0) {
spa_load_failed(spa, "spa_ld_unflushed_txgs(): "
"failed at dmu_read(obj=%llu) [error %d]",
(u_longlong_t)object, error);
return (error);
}
ms->ms_unflushed_txg = entry.msp_unflushed_txg;
if (ms->ms_unflushed_txg != 0) {
mutex_enter(&spa->spa_flushed_ms_lock);
avl_add(&spa->spa_metaslabs_by_flushed, ms);
mutex_exit(&spa->spa_flushed_ms_lock);
}
}
return (0);
}
/*
* Read all the log space map entries into their respective
* metaslab unflushed trees and keep them sorted by TXG in the
* SPA's metadata. In addition, setup all the metadata for the
* memory and the block heuristics.
*/
int
spa_ld_log_spacemaps(spa_t *spa)
{
int error;
spa_log_sm_set_blocklimit(spa);
for (uint64_t c = 0; c < spa->spa_root_vdev->vdev_children; c++) {
vdev_t *vd = spa->spa_root_vdev->vdev_child[c];
error = spa_ld_unflushed_txgs(vd);
if (error != 0)
return (error);
}
error = spa_ld_log_sm_metadata(spa);
if (error != 0)
return (error);
/*
* Note: we don't actually expect anything to change at this point
* but we grab the config lock so we don't fail any assertions
* when using vdev_lookup_top().
*/
spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
error = spa_ld_log_sm_data(spa);
spa_config_exit(spa, SCL_CONFIG, FTAG);
return (error);
}
/* BEGIN CSTYLED */
ZFS_MODULE_PARAM(zfs, zfs_, unflushed_max_mem_amt, ULONG, ZMOD_RW,
"Specific hard-limit in memory that ZFS allows to be used for "
"unflushed changes");
Log Spacemap Project = Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent db587941c5ff6dea01932bb78f70db63cf7f38ba Update vdev_is_spacemap_addressable() for new spacemap encoding 419ba5914552c6185afbe1dd17b3ed4b0d526547 Simplify spa_sync by breaking it up to smaller functions 8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834 Factor metaslab_load_wait() in metaslab_load() b194fab0fb6caad18711abccaff3c69ad8b3f6d3 Rename range_tree_verify to range_tree_verify_not_present df72b8bebe0ebac0b20e0750984bad182cb6564a Change target size of metaslabs from 256GB to 16GB c853f382db731e15a87512f4ef1101d14d778a55 zdb -L should skip leak detection altogether 21e7cf5da89f55ce98ec1115726b150e19eefe89 vs_alloc can underflow in L2ARC vdevs 7558997d2f808368867ca7e5234e5793446e8f3f Simplify log vdev removal code 6c926f426a26ffb6d7d8e563e33fc176164175cb Get rid of space_map_update() for ms_synced_length 425d3237ee88abc53d8522a7139c926d278b4b7f Introduce auxiliary metaslab histograms 928e8ad47d3478a3d5d01f0dd6ae74a9371af65e Error path in metaslab_load_impl() forgets to drop ms_sync_lock 8eef997679ba54547f7d361553d21b3291f41ae7 = References Background, Motivation, and Internals of the Feature - OpenZFS 2017 Presentation: youtu.be/jj2IxRkl5bQ - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project Flushing Algorithm Internals & Performance Results (Illumos Specific) - Blogpost: sdimitro.github.io/post/zfs-lsm-flushing/ - OpenZFS 2018 Presentation: youtu.be/x6D2dHRjkxw - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm Upstream Delphix Issues: DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320 DLPX-63385 Reviewed-by: Sean Eric Fagan <sef@ixsystems.com> Reviewed-by: Matt Ahrens <matt@delphix.com> Reviewed-by: George Wilson <gwilson@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com> Closes #8442
2019-07-16 20:11:49 +03:00
ZFS_MODULE_PARAM(zfs, zfs_, unflushed_max_mem_ppm, ULONG, ZMOD_RW,
"Percentage of the overall system memory that ZFS allows to be "
"used for unflushed changes (value is calculated over 1000000 for "
"finer granularity)");
Log Spacemap Project = Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent db587941c5ff6dea01932bb78f70db63cf7f38ba Update vdev_is_spacemap_addressable() for new spacemap encoding 419ba5914552c6185afbe1dd17b3ed4b0d526547 Simplify spa_sync by breaking it up to smaller functions 8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834 Factor metaslab_load_wait() in metaslab_load() b194fab0fb6caad18711abccaff3c69ad8b3f6d3 Rename range_tree_verify to range_tree_verify_not_present df72b8bebe0ebac0b20e0750984bad182cb6564a Change target size of metaslabs from 256GB to 16GB c853f382db731e15a87512f4ef1101d14d778a55 zdb -L should skip leak detection altogether 21e7cf5da89f55ce98ec1115726b150e19eefe89 vs_alloc can underflow in L2ARC vdevs 7558997d2f808368867ca7e5234e5793446e8f3f Simplify log vdev removal code 6c926f426a26ffb6d7d8e563e33fc176164175cb Get rid of space_map_update() for ms_synced_length 425d3237ee88abc53d8522a7139c926d278b4b7f Introduce auxiliary metaslab histograms 928e8ad47d3478a3d5d01f0dd6ae74a9371af65e Error path in metaslab_load_impl() forgets to drop ms_sync_lock 8eef997679ba54547f7d361553d21b3291f41ae7 = References Background, Motivation, and Internals of the Feature - OpenZFS 2017 Presentation: youtu.be/jj2IxRkl5bQ - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project Flushing Algorithm Internals & Performance Results (Illumos Specific) - Blogpost: sdimitro.github.io/post/zfs-lsm-flushing/ - OpenZFS 2018 Presentation: youtu.be/x6D2dHRjkxw - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm Upstream Delphix Issues: DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320 DLPX-63385 Reviewed-by: Sean Eric Fagan <sef@ixsystems.com> Reviewed-by: Matt Ahrens <matt@delphix.com> Reviewed-by: George Wilson <gwilson@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com> Closes #8442
2019-07-16 20:11:49 +03:00
ZFS_MODULE_PARAM(zfs, zfs_, unflushed_log_block_max, ULONG, ZMOD_RW,
"Hard limit (upper-bound) in the size of the space map log "
"in terms of blocks.");
Log Spacemap Project = Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent db587941c5ff6dea01932bb78f70db63cf7f38ba Update vdev_is_spacemap_addressable() for new spacemap encoding 419ba5914552c6185afbe1dd17b3ed4b0d526547 Simplify spa_sync by breaking it up to smaller functions 8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834 Factor metaslab_load_wait() in metaslab_load() b194fab0fb6caad18711abccaff3c69ad8b3f6d3 Rename range_tree_verify to range_tree_verify_not_present df72b8bebe0ebac0b20e0750984bad182cb6564a Change target size of metaslabs from 256GB to 16GB c853f382db731e15a87512f4ef1101d14d778a55 zdb -L should skip leak detection altogether 21e7cf5da89f55ce98ec1115726b150e19eefe89 vs_alloc can underflow in L2ARC vdevs 7558997d2f808368867ca7e5234e5793446e8f3f Simplify log vdev removal code 6c926f426a26ffb6d7d8e563e33fc176164175cb Get rid of space_map_update() for ms_synced_length 425d3237ee88abc53d8522a7139c926d278b4b7f Introduce auxiliary metaslab histograms 928e8ad47d3478a3d5d01f0dd6ae74a9371af65e Error path in metaslab_load_impl() forgets to drop ms_sync_lock 8eef997679ba54547f7d361553d21b3291f41ae7 = References Background, Motivation, and Internals of the Feature - OpenZFS 2017 Presentation: youtu.be/jj2IxRkl5bQ - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project Flushing Algorithm Internals & Performance Results (Illumos Specific) - Blogpost: sdimitro.github.io/post/zfs-lsm-flushing/ - OpenZFS 2018 Presentation: youtu.be/x6D2dHRjkxw - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm Upstream Delphix Issues: DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320 DLPX-63385 Reviewed-by: Sean Eric Fagan <sef@ixsystems.com> Reviewed-by: Matt Ahrens <matt@delphix.com> Reviewed-by: George Wilson <gwilson@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com> Closes #8442
2019-07-16 20:11:49 +03:00
ZFS_MODULE_PARAM(zfs, zfs_, unflushed_log_block_min, ULONG, ZMOD_RW,
"Lower-bound limit for the maximum amount of blocks allowed in "
"log spacemap (see zfs_unflushed_log_block_max)");
Log Spacemap Project = Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent db587941c5ff6dea01932bb78f70db63cf7f38ba Update vdev_is_spacemap_addressable() for new spacemap encoding 419ba5914552c6185afbe1dd17b3ed4b0d526547 Simplify spa_sync by breaking it up to smaller functions 8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834 Factor metaslab_load_wait() in metaslab_load() b194fab0fb6caad18711abccaff3c69ad8b3f6d3 Rename range_tree_verify to range_tree_verify_not_present df72b8bebe0ebac0b20e0750984bad182cb6564a Change target size of metaslabs from 256GB to 16GB c853f382db731e15a87512f4ef1101d14d778a55 zdb -L should skip leak detection altogether 21e7cf5da89f55ce98ec1115726b150e19eefe89 vs_alloc can underflow in L2ARC vdevs 7558997d2f808368867ca7e5234e5793446e8f3f Simplify log vdev removal code 6c926f426a26ffb6d7d8e563e33fc176164175cb Get rid of space_map_update() for ms_synced_length 425d3237ee88abc53d8522a7139c926d278b4b7f Introduce auxiliary metaslab histograms 928e8ad47d3478a3d5d01f0dd6ae74a9371af65e Error path in metaslab_load_impl() forgets to drop ms_sync_lock 8eef997679ba54547f7d361553d21b3291f41ae7 = References Background, Motivation, and Internals of the Feature - OpenZFS 2017 Presentation: youtu.be/jj2IxRkl5bQ - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project Flushing Algorithm Internals & Performance Results (Illumos Specific) - Blogpost: sdimitro.github.io/post/zfs-lsm-flushing/ - OpenZFS 2018 Presentation: youtu.be/x6D2dHRjkxw - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm Upstream Delphix Issues: DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320 DLPX-63385 Reviewed-by: Sean Eric Fagan <sef@ixsystems.com> Reviewed-by: Matt Ahrens <matt@delphix.com> Reviewed-by: George Wilson <gwilson@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com> Closes #8442
2019-07-16 20:11:49 +03:00
ZFS_MODULE_PARAM(zfs, zfs_, unflushed_log_block_pct, ULONG, ZMOD_RW,
"Tunable used to determine the number of blocks that can be used for "
"the spacemap log, expressed as a percentage of the total number of "
"metaslabs in the pool (e.g. 400 means the number of log blocks is "
"capped at 4 times the number of metaslabs)");
Log Spacemap Project = Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent db587941c5ff6dea01932bb78f70db63cf7f38ba Update vdev_is_spacemap_addressable() for new spacemap encoding 419ba5914552c6185afbe1dd17b3ed4b0d526547 Simplify spa_sync by breaking it up to smaller functions 8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834 Factor metaslab_load_wait() in metaslab_load() b194fab0fb6caad18711abccaff3c69ad8b3f6d3 Rename range_tree_verify to range_tree_verify_not_present df72b8bebe0ebac0b20e0750984bad182cb6564a Change target size of metaslabs from 256GB to 16GB c853f382db731e15a87512f4ef1101d14d778a55 zdb -L should skip leak detection altogether 21e7cf5da89f55ce98ec1115726b150e19eefe89 vs_alloc can underflow in L2ARC vdevs 7558997d2f808368867ca7e5234e5793446e8f3f Simplify log vdev removal code 6c926f426a26ffb6d7d8e563e33fc176164175cb Get rid of space_map_update() for ms_synced_length 425d3237ee88abc53d8522a7139c926d278b4b7f Introduce auxiliary metaslab histograms 928e8ad47d3478a3d5d01f0dd6ae74a9371af65e Error path in metaslab_load_impl() forgets to drop ms_sync_lock 8eef997679ba54547f7d361553d21b3291f41ae7 = References Background, Motivation, and Internals of the Feature - OpenZFS 2017 Presentation: youtu.be/jj2IxRkl5bQ - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project Flushing Algorithm Internals & Performance Results (Illumos Specific) - Blogpost: sdimitro.github.io/post/zfs-lsm-flushing/ - OpenZFS 2018 Presentation: youtu.be/x6D2dHRjkxw - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm Upstream Delphix Issues: DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320 DLPX-63385 Reviewed-by: Sean Eric Fagan <sef@ixsystems.com> Reviewed-by: Matt Ahrens <matt@delphix.com> Reviewed-by: George Wilson <gwilson@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com> Closes #8442
2019-07-16 20:11:49 +03:00
ZFS_MODULE_PARAM(zfs, zfs_, max_log_walking, ULONG, ZMOD_RW,
"The number of past TXGs that the flushing algorithm of the log "
"spacemap feature uses to estimate incoming log blocks");
Log Spacemap Project = Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent db587941c5ff6dea01932bb78f70db63cf7f38ba Update vdev_is_spacemap_addressable() for new spacemap encoding 419ba5914552c6185afbe1dd17b3ed4b0d526547 Simplify spa_sync by breaking it up to smaller functions 8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834 Factor metaslab_load_wait() in metaslab_load() b194fab0fb6caad18711abccaff3c69ad8b3f6d3 Rename range_tree_verify to range_tree_verify_not_present df72b8bebe0ebac0b20e0750984bad182cb6564a Change target size of metaslabs from 256GB to 16GB c853f382db731e15a87512f4ef1101d14d778a55 zdb -L should skip leak detection altogether 21e7cf5da89f55ce98ec1115726b150e19eefe89 vs_alloc can underflow in L2ARC vdevs 7558997d2f808368867ca7e5234e5793446e8f3f Simplify log vdev removal code 6c926f426a26ffb6d7d8e563e33fc176164175cb Get rid of space_map_update() for ms_synced_length 425d3237ee88abc53d8522a7139c926d278b4b7f Introduce auxiliary metaslab histograms 928e8ad47d3478a3d5d01f0dd6ae74a9371af65e Error path in metaslab_load_impl() forgets to drop ms_sync_lock 8eef997679ba54547f7d361553d21b3291f41ae7 = References Background, Motivation, and Internals of the Feature - OpenZFS 2017 Presentation: youtu.be/jj2IxRkl5bQ - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project Flushing Algorithm Internals & Performance Results (Illumos Specific) - Blogpost: sdimitro.github.io/post/zfs-lsm-flushing/ - OpenZFS 2018 Presentation: youtu.be/x6D2dHRjkxw - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm Upstream Delphix Issues: DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320 DLPX-63385 Reviewed-by: Sean Eric Fagan <sef@ixsystems.com> Reviewed-by: Matt Ahrens <matt@delphix.com> Reviewed-by: George Wilson <gwilson@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com> Closes #8442
2019-07-16 20:11:49 +03:00
ZFS_MODULE_PARAM(zfs, zfs_, max_logsm_summary_length, ULONG, ZMOD_RW,
"Maximum number of rows allowed in the summary of the spacemap log");
Log Spacemap Project = Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent db587941c5ff6dea01932bb78f70db63cf7f38ba Update vdev_is_spacemap_addressable() for new spacemap encoding 419ba5914552c6185afbe1dd17b3ed4b0d526547 Simplify spa_sync by breaking it up to smaller functions 8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834 Factor metaslab_load_wait() in metaslab_load() b194fab0fb6caad18711abccaff3c69ad8b3f6d3 Rename range_tree_verify to range_tree_verify_not_present df72b8bebe0ebac0b20e0750984bad182cb6564a Change target size of metaslabs from 256GB to 16GB c853f382db731e15a87512f4ef1101d14d778a55 zdb -L should skip leak detection altogether 21e7cf5da89f55ce98ec1115726b150e19eefe89 vs_alloc can underflow in L2ARC vdevs 7558997d2f808368867ca7e5234e5793446e8f3f Simplify log vdev removal code 6c926f426a26ffb6d7d8e563e33fc176164175cb Get rid of space_map_update() for ms_synced_length 425d3237ee88abc53d8522a7139c926d278b4b7f Introduce auxiliary metaslab histograms 928e8ad47d3478a3d5d01f0dd6ae74a9371af65e Error path in metaslab_load_impl() forgets to drop ms_sync_lock 8eef997679ba54547f7d361553d21b3291f41ae7 = References Background, Motivation, and Internals of the Feature - OpenZFS 2017 Presentation: youtu.be/jj2IxRkl5bQ - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project Flushing Algorithm Internals & Performance Results (Illumos Specific) - Blogpost: sdimitro.github.io/post/zfs-lsm-flushing/ - OpenZFS 2018 Presentation: youtu.be/x6D2dHRjkxw - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm Upstream Delphix Issues: DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320 DLPX-63385 Reviewed-by: Sean Eric Fagan <sef@ixsystems.com> Reviewed-by: Matt Ahrens <matt@delphix.com> Reviewed-by: George Wilson <gwilson@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com> Closes #8442
2019-07-16 20:11:49 +03:00
ZFS_MODULE_PARAM(zfs, zfs_, min_metaslabs_to_flush, ULONG, ZMOD_RW,
"Minimum number of metaslabs to flush per dirty TXG");
Log Spacemap Project = Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent db587941c5ff6dea01932bb78f70db63cf7f38ba Update vdev_is_spacemap_addressable() for new spacemap encoding 419ba5914552c6185afbe1dd17b3ed4b0d526547 Simplify spa_sync by breaking it up to smaller functions 8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834 Factor metaslab_load_wait() in metaslab_load() b194fab0fb6caad18711abccaff3c69ad8b3f6d3 Rename range_tree_verify to range_tree_verify_not_present df72b8bebe0ebac0b20e0750984bad182cb6564a Change target size of metaslabs from 256GB to 16GB c853f382db731e15a87512f4ef1101d14d778a55 zdb -L should skip leak detection altogether 21e7cf5da89f55ce98ec1115726b150e19eefe89 vs_alloc can underflow in L2ARC vdevs 7558997d2f808368867ca7e5234e5793446e8f3f Simplify log vdev removal code 6c926f426a26ffb6d7d8e563e33fc176164175cb Get rid of space_map_update() for ms_synced_length 425d3237ee88abc53d8522a7139c926d278b4b7f Introduce auxiliary metaslab histograms 928e8ad47d3478a3d5d01f0dd6ae74a9371af65e Error path in metaslab_load_impl() forgets to drop ms_sync_lock 8eef997679ba54547f7d361553d21b3291f41ae7 = References Background, Motivation, and Internals of the Feature - OpenZFS 2017 Presentation: youtu.be/jj2IxRkl5bQ - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project Flushing Algorithm Internals & Performance Results (Illumos Specific) - Blogpost: sdimitro.github.io/post/zfs-lsm-flushing/ - OpenZFS 2018 Presentation: youtu.be/x6D2dHRjkxw - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm Upstream Delphix Issues: DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320 DLPX-63385 Reviewed-by: Sean Eric Fagan <sef@ixsystems.com> Reviewed-by: Matt Ahrens <matt@delphix.com> Reviewed-by: George Wilson <gwilson@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com> Closes #8442
2019-07-16 20:11:49 +03:00
ZFS_MODULE_PARAM(zfs, zfs_, keep_log_spacemaps_at_export, INT, ZMOD_RW,
"Prevent the log spacemaps from being flushed and destroyed "
"during pool export/destroy");
Log Spacemap Project = Motivation At Delphix we've seen a lot of customer systems where fragmentation is over 75% and random writes take a performance hit because a lot of time is spend on I/Os that update on-disk space accounting metadata. Specifically, we seen cases where 20% to 40% of sync time is spend after sync pass 1 and ~30% of the I/Os on the system is spent updating spacemaps. The problem is that these pools have existed long enough that we've touched almost every metaslab at least once, and random writes scatter frees across all metaslabs every TXG, thus appending to their spacemaps and resulting in many I/Os. To give an example, assuming that every VDEV has 200 metaslabs and our writes fit within a single spacemap block (generally 4K) we have 200 I/Os. Then if we assume 2 levels of indirection, we need 400 additional I/Os and since we are talking about metadata for which we keep 2 extra copies for redundancy we need to triple that number, leading to a total of 1800 I/Os per VDEV every TXG. We could try and decrease the number of metaslabs so we have less I/Os per TXG but then each metaslab would cover a wider range on disk and thus would take more time to be loaded in memory from disk. In addition, after it's loaded, it's range tree would consume more memory. Another idea would be to just increase the spacemap block size which would allow us to fit more entries within an I/O block resulting in fewer I/Os per metaslab and a speedup in loading time. The problem is still that we don't deal with the number of I/Os going up as the number of metaslabs is increasing and the fact is that we generally write a lot to a few metaslabs and a little to the rest of them. Thus, just increasing the block size would actually waste bandwidth because we won't be utilizing our bigger block size. = About this patch This patch introduces the Log Spacemap project which provides the solution to the above problem while taking into account all the aforementioned tradeoffs. The details on how it achieves that can be found in the references sections below and in the code (see Big Theory Statement in spa_log_spacemap.c). Even though the change is fairly constraint within the metaslab and lower-level SPA codepaths, there is a side-change that is user-facing. The change is that VDEV IDs from VDEV holes will no longer be reused. To give some background and reasoning for this, when a log device is removed and its VDEV structure was replaced with a hole (or was compacted; if at the end of the vdev array), its vdev_id could be reused by devices added after that. Now with the pool-wide space maps recording the vdev ID, this behavior can cause problems (e.g. is this entry referring to a segment in the new vdev or the removed log?). Thus, to simplify things the ID reuse behavior is gone and now vdev IDs for top-level vdevs are truly unique within a pool. = Testing The illumos implementation of this feature has been used internally for a year and has been in production for ~6 months. For this patch specifically there don't seem to be any regressions introduced to ZTS and I have been running zloop for a week without any related problems. = Performance Analysis (Linux Specific) All performance results and analysis for illumos can be found in the links of the references. Redoing the same experiments in Linux gave similar results. Below are the specifics of the Linux run. After the pool reached stable state the percentage of the time spent in pass 1 per TXG was 64% on average for the stock bits while the log spacemap bits stayed at 95% during the experiment (graph: sdimitro.github.io/img/linux-lsm/PercOfSyncInPassOne.png). Sync times per TXG were 37.6 seconds on average for the stock bits and 22.7 seconds for the log spacemap bits (related graph: sdimitro.github.io/img/linux-lsm/SyncTimePerTXG.png). As a result the log spacemap bits were able to push more TXGs, which is also the reason why all graphs quantified per TXG have more entries for the log spacemap bits. Another interesting aspect in terms of txg syncs is that the stock bits had 22% of their TXGs reach sync pass 7, 55% reach sync pass 8, and 20% reach 9. The log space map bits reached sync pass 4 in 79% of their TXGs, sync pass 7 in 19%, and sync pass 8 at 1%. This emphasizes the fact that not only we spend less time on metadata but we also iterate less times to convergence in spa_sync() dirtying objects. [related graphs: stock- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGStock.png lsm- sdimitro.github.io/img/linux-lsm/NumberOfPassesPerTXGLSM.png] Finally, the improvement in IOPs that the userland gains from the change is approximately 40%. There is a consistent win in IOPS as you can see from the graphs below but the absolute amount of improvement that the log spacemap gives varies within each minute interval. sdimitro.github.io/img/linux-lsm/StockVsLog3Days.png sdimitro.github.io/img/linux-lsm/StockVsLog10Hours.png = Porting to Other Platforms For people that want to port this commit to other platforms below is a list of ZoL commits that this patch depends on: Make zdb results for checkpoint tests consistent db587941c5ff6dea01932bb78f70db63cf7f38ba Update vdev_is_spacemap_addressable() for new spacemap encoding 419ba5914552c6185afbe1dd17b3ed4b0d526547 Simplify spa_sync by breaking it up to smaller functions 8dc2197b7b1e4d7ebc1420ea30e51c6541f1d834 Factor metaslab_load_wait() in metaslab_load() b194fab0fb6caad18711abccaff3c69ad8b3f6d3 Rename range_tree_verify to range_tree_verify_not_present df72b8bebe0ebac0b20e0750984bad182cb6564a Change target size of metaslabs from 256GB to 16GB c853f382db731e15a87512f4ef1101d14d778a55 zdb -L should skip leak detection altogether 21e7cf5da89f55ce98ec1115726b150e19eefe89 vs_alloc can underflow in L2ARC vdevs 7558997d2f808368867ca7e5234e5793446e8f3f Simplify log vdev removal code 6c926f426a26ffb6d7d8e563e33fc176164175cb Get rid of space_map_update() for ms_synced_length 425d3237ee88abc53d8522a7139c926d278b4b7f Introduce auxiliary metaslab histograms 928e8ad47d3478a3d5d01f0dd6ae74a9371af65e Error path in metaslab_load_impl() forgets to drop ms_sync_lock 8eef997679ba54547f7d361553d21b3291f41ae7 = References Background, Motivation, and Internals of the Feature - OpenZFS 2017 Presentation: youtu.be/jj2IxRkl5bQ - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemaps-project Flushing Algorithm Internals & Performance Results (Illumos Specific) - Blogpost: sdimitro.github.io/post/zfs-lsm-flushing/ - OpenZFS 2018 Presentation: youtu.be/x6D2dHRjkxw - Slides: slideshare.net/SerapheimNikolaosDim/zfs-log-spacemap-flushing-algorithm Upstream Delphix Issues: DLPX-51539, DLPX-59659, DLPX-57783, DLPX-61438, DLPX-41227, DLPX-59320 DLPX-63385 Reviewed-by: Sean Eric Fagan <sef@ixsystems.com> Reviewed-by: Matt Ahrens <matt@delphix.com> Reviewed-by: George Wilson <gwilson@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Serapheim Dimitropoulos <serapheim@delphix.com> Closes #8442
2019-07-16 20:11:49 +03:00
/* END CSTYLED */