= Motivation
While dealing with another performance issue (see 126118f) we noticed
that we spend a lot of time in various places in the kernel when
constructing long nvlists. The problem is that when an nvlist is created
with the NV_UNIQUE_NAME set (which is the case most of the time), we do
a linear search through the whole list to ensure uniqueness for every
entry we add.
An example of the above scenario can be seen in the following
flamegraph, where more than have the time of the zfsdev_ioctl() is spent
on constructing nvlists. Flamegraph:
https://sdimitro.github.io/img/flame/sdimitro_snap_unmount3.svg
Adding a table to speed up lookups will help situations where we just
construct an nvlist (like the scenario above), in addition to regular
lookups and removals.
= What this patch does
In this diff we've implemented a hash-table on top of the nvlist code
that converts most nvlist operations from O(# number of entries) to
O(1)* (the start is for amortized time as the hash-table grows and
shrinks depending on the # of entries - plain lookup is strictly O(1)).
= Performance Analysis
To analyze the performance improvement I just used the setup from the
snapshot deletion issue mentioned above in the Motivation section.
Basically I created 10K filesystems with one snapshot each and then I
just used the API of libZFS_Core to pass down an nvlist of all the
snapshots to have them deleted. The reason I used my own driver program
was to have clean performance results of what actually happens in the
kernel. The flamegraphs and wall clock times mentioned below were
gathered from the start to the end of the driver program's run. Between
trials the testpool used was completely destroyed, the system was
rebooted and the testpool was completely recreated. The reason for this
dance was to get consistent results.
== Results (before patch):
=== Sampling Flamegraphs
[Trial 1] https://sdimitro.github.io/img/flame/DLPX-53417/trial-A.svg
[Trial 2] https://sdimitro.github.io/img/flame/DLPX-53417/trial-A2.svg
[Trial 3] https://sdimitro.github.io/img/flame/DLPX-53417/trial-A3.svg
=== Wall clock times (in seconds)
```
[Trial 4]
real 5.3
user 0.4
sys 2.3
[Trial 5]
real 8.2
user 0.4
sys 2.4
[Trial 6]
real 6.0
user 0.5
sys 2.3
```
== Results (after patch):
=== Sampling Flamegraphs
[Trial 1] https://sdimitro.github.io/img/flame/DLPX-53417/trial-Ae.svg
[Trial 2] https://sdimitro.github.io/img/flame/DLPX-53417/trial-A2e.svg
[Trial 3] https://sdimitro.github.io/img/flame/DLPX-53417/trial-A3e.svg
=== Wall clock times (in seconds)
```
[Trial 4]
real 4.9
user 0.0
sys 0.9
[Trial 5]
real 3.8
user 0.0
sys 0.9
[Trial 6]
real 3.6
user 0.0
sys 0.9
```
== Analysis
The results between the trials are consistent so in this sections I will
only talk about the flamegraph results from trial-1 and the wall-clock
results from trial-4.
From trial-1 we can see that zfs_dev_ioctl() goes from 2,331 to 996
samples counts. Specifically, the samples from fnvlist_add_nvlist() and
spa_history_log_nvl() are almost gone (~500 & ~800 to 5 & 5 samples),
leaving zfs_ioc_destroy_snaps() to dominate most samples from
zfs_dev_ioctl().
From trial-4 we see that the user time dropped to 0 secods. I believe
the consistent 0.4 seconds before my patch was applied was due to my
driver program constructing the long nvlist of snapshots so it can pass
it to the kernel. As for the system time, the effect there is more clear
(2.3 down to 0.9 seconds).
Porting Notes:
* DATA_TYPE_DONTCARE case added to switch in fm_nvprintr() and
zpool_do_events_nvprint().
Authored by: Serapheim Dimitropoulos <serapheim.dimitro@delphix.com>
Reviewed by: Matt Ahrens <matt@delphix.com>
Reviewed by: Sebastien Roy <sebastien.roy@delphix.com>
Approved by: Robert Mustacchi <rm@joyent.com>
Ported-by: Brian Behlendorf <behlendorf1@llnl.gov>
OpenZFS-issue: https://www.illumos.org/issues/9580
OpenZFS-commit: https://github.com/openzfs/openzfs/commit/b5eca7b1Closes#7748
6b64382b17
