mirror_zfs/man/man5/spl-module-parameters.5
Brian Behlendorf 3018bffa9b Refine slab cache sizing
This change is designed to improve the memory utilization of
slabs by more carefully setting their size.  The way the code
currently works is problematic for slabs which contain large
objects (>1MB).  This is due to slabs being unconditionally
rounded up to a power of two which may result in unused space
at the end of the slab.

The reason the existing code rounds up every slab is because it
assumes it will backed by the buddy allocator.  Since the buddy
allocator can only performs power of two allocations this is
desirable because it avoids wasting any space.  However, this
logic breaks down if slab is backed by vmalloc() which operates
at a page level granularity.  In this case, the optimal thing to
do is calculate the minimum required slab size given certain
constraints (object size, alignment, objects/slab, etc).

Therefore, this patch reworks the spl_slab_size() function so
that it sizes KMC_KMEM slabs differently than KMC_VMEM slabs.
KMC_KMEM slabs are rounded up to the nearest power of two, and
KMC_VMEM slabs are allowed to be the minimum required size.

This change also reduces the default number of objects per slab.
This reduces how much memory a single cache object can pin, which
can result in significant memory saving for highly fragmented
caches.  But depending on the workload it may result in slabs
being allocated and freed more frequently.  In practice, this
has been shown to be a better default for most workloads.

Also the maximum slab size has been reduced to 4MB on 32-bit
systems.  Due to the limited virtual address space it's critical
the we be as frugal as possible.  A limit of 4M still lets us
reasonably comfortably allocate a limited number of 1MB objects.

Finally, the kmem:slab_small and kmem:slab_large SPLAT tests
were extended to provide better test coverage of various object
sizes and alignments.  Caches are created with random parameters
and their basic functionality is verified by allocating several
slabs worth of objects.

Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>
2015-01-16 13:55:09 -08:00

252 lines
7.9 KiB
Groff

'\" te
.\"
.\" Copyright 2013 Turbo Fredriksson <turbo@bayour.com>. All rights reserved.
.\"
.TH SPL-MODULE-PARAMETERS 5 "Nov 18, 2013"
.SH NAME
spl\-module\-parameters \- SPL module parameters
.SH DESCRIPTION
.sp
.LP
Description of the different parameters to the SPL module.
.SS "Module parameters"
.sp
.LP
.sp
.ne 2
.na
\fBspl_kmem_cache_expire\fR (uint)
.ad
.RS 12n
Cache expiration is part of default Illumos cache behavior. The idea is
that objects in magazines which have not been recently accessed should be
returned to the slabs periodically. This is known as cache aging and
when enabled objects will be typically returned after 15 seconds.
.sp
On the other hand Linux slabs are designed to never move objects back to
the slabs unless there is memory pressure. This is possible because under
Linux the cache will be notified when memory is low and objects can be
released.
.sp
By default only the Linux method is enabled. It has been shown to improve
responsiveness on low memory systems and not negatively impact the performance
of systems with more memory. This policy may be changed by setting the
\fBspl_kmem_cache_expire\fR bit mask as follows, both policies may be enabled
concurrently.
.sp
0x01 - Aging (Illumos), 0x02 - Low memory (Linux)
.sp
Default value: \fB0x02\fR
.RE
.sp
.ne 2
.na
\fBspl_kmem_cache_reclaim\fR (uint)
.ad
.RS 12n
When this is set it prevents Linux from being able to rapidly reclaim all the
memory held by the kmem caches. This may be useful in circumstances where
it's preferable that Linux reclaim memory from some other subsystem first.
Setting this will increase the likelihood out of memory events on a memory
constrained system.
.sp
Default value: \fB0\fR
.RE
.sp
.ne 2
.na
\fBspl_kmem_cache_obj_per_slab\fR (uint)
.ad
.RS 12n
The preferred number of objects per slab in the cache. In general, a larger
value will increase the caches memory footprint while decreasing the time
required to perform an allocation. Conversely, a smaller value will minimize
the footprint and improve cache reclaim time but individual allocations may
take longer.
.sp
Default value: \fB8\fR
.RE
.sp
.ne 2
.na
\fBspl_kmem_cache_obj_per_slab_min\fR (uint)
.ad
.RS 12n
The minimum number of objects allowed per slab. Normally slabs will contain
\fBspl_kmem_cache_obj_per_slab\fR objects but for caches that contain very
large objects it's desirable to only have a few, or even just one, object per
slab.
.sp
Default value: \fB1\fR
.RE
.sp
.ne 2
.na
\fBspl_kmem_cache_max_size\fR (uint)
.ad
.RS 12n
The maximum size of a kmem cache slab in MiB. This effectively limits
the maximum cache object size to \fBspl_kmem_cache_max_size\fR /
\fBspl_kmem_cache_obj_per_slab\fR. Caches may not be created with
object sized larger than this limit.
.sp
Default value: \fB32 (64-bit) or 4 (32-bit)\fR
.RE
.sp
.ne 2
.na
\fBspl_kmem_cache_slab_limit\fR (uint)
.ad
.RS 12n
For small objects the Linux slab allocator should be used to make the most
efficient use of the memory. However, large objects are not supported by
the Linux slab and therefore the SPL implementation is preferred. This
value is used to determine the cutoff between a small and large object.
.sp
Objects of \fBspl_kmem_cache_slab_limit\fR or smaller will be allocated
using the Linux slab allocator, large objects use the SPL allocator. A
cutoff of 16K was determined to be optimal for architectures using 4K pages.
.sp
Default value: \fB16,384\fR
.RE
.sp
.ne 2
.na
\fBspl_kmem_cache_kmem_limit\fR (uint)
.ad
.RS 12n
Depending on the size of a cache object it may be backed by kmalloc()'d
or vmalloc()'d memory. This is because the size of the required allocation
greatly impacts the best way to allocate the memory.
.sp
When objects are small and only a small number of memory pages need to be
allocated, ideally just one, then kmalloc() is very efficient. However,
when allocating multiple pages with kmalloc() it gets increasingly expensive
because the pages must be physically contiguous.
.sp
For this reason we shift to vmalloc() for slabs of large objects which
which removes the need for contiguous pages. We cannot use vmalloc() in
all cases because there is significant locking overhead involved. This
function takes a single global lock over the entire virtual address range
which serializes all allocations. Using slightly different allocation
functions for small and large objects allows us to handle a wide range of
object sizes.
.sh
The \fBspl_kmem_cache_kmem_limit\fR value is used to determine this cutoff
size. One quarter the PAGE_SIZE is used as the default value because
\fBspl_kmem_cache_obj_per_slab\fR defaults to 16. This means that at
most we will need to allocate four contiguous pages.
.sp
Default value: \fBPAGE_SIZE/4\fR
.RE
.sp
.ne 2
.na
\fBspl_kmem_alloc_warn\fR (uint)
.ad
.RS 12n
As a general rule kmem_alloc() allocations should be small, preferably
just a few pages since they must by physically contiguous. Therefore, a
rate limited warning will be printed to the console for any kmem_alloc()
which exceeds a reasonable threshold.
.sp
The default warning threshold is set to eight pages but capped at 32K to
accommodate systems using large pages. This value was selected to be small
enough to ensure the largest allocations are quickly noticed and fixed.
But large enough to avoid logging any warnings when a allocation size is
larger than optimal but not a serious concern. Since this value is tunable,
developers are encouraged to set it lower when testing so any new largish
allocations are quickly caught. These warnings may be disabled by setting
the threshold to zero.
.sp
Default value: \fB32,768\fR
.RE
.sp
.ne 2
.na
\fBspl_kmem_alloc_max\fR (uint)
.ad
.RS 12n
Large kmem_alloc() allocations will fail if they exceed KMALLOC_MAX_SIZE.
Allocations which are marginally smaller than this limit may succeed but
should still be avoided due to the expense of locating a contiguous range
of free pages. Therefore, a maximum kmem size with reasonable safely
margin of 4x is set. Kmem_alloc() allocations larger than this maximum
will quickly fail. Vmem_alloc() allocations less than or equal to this
value will use kmalloc(), but shift to vmalloc() when exceeding this value.
.sp
Default value: \fBKMALLOC_MAX_SIZE/4\fR
.RE
.sp
.ne 2
.na
\fBspl_kmem_cache_magazine_size\fR (uint)
.ad
.RS 12n
Cache magazines are an optimization designed to minimize the cost of
allocating memory. They do this by keeping a per-cpu cache of recently
freed objects, which can then be reallocated without taking a lock. This
can improve performance on highly contended caches. However, because
objects in magazines will prevent otherwise empty slabs from being
immediately released this may not be ideal for low memory machines.
.sp
For this reason \fBspl_kmem_cache_magazine_size\fR can be used to set a
maximum magazine size. When this value is set to 0 the magazine size will
be automatically determined based on the object size. Otherwise magazines
will be limited to 2-256 objects per magazine (i.e per cpu). Magazines
may never be entirely disabled in this implementation.
.sp
Default value: \fB0\fR
.RE
.sp
.ne 2
.na
\fBspl_hostid\fR (ulong)
.ad
.RS 12n
The system hostid, when set this can be used to uniquely identify a system.
By default this value is set to zero which indicates the hostid is disabled.
It can be explicitly enabled by placing a unique non-zero value in
\fB/etc/hostid/\fR.
.sp
Default value: \fB0\fR
.RE
.sp
.ne 2
.na
\fBspl_hostid_path\fR (charp)
.ad
.RS 12n
The expected path to locate the system hostid when specified. This value
may be overridden for non-standard configurations.
.sp
Default value: \fB/etc/hostid\fR
.RE
.sp
.ne 2
.na
\fBspl_taskq_thread_bind\fR (int)
.ad
.RS 12n
Bind taskq threads to specific CPUs. When enabled all taskq threads will
be distributed evenly over the available CPUs. By default, this behavior
is disabled to allow the Linux scheduler the maximum flexibility to determine
where a thread should run.
.sp
Default value: \fB0\fR
.RE