Compact dbuf/buf hashes and lock arrays. #12289
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
With default dbuf cache size of 1/32 of ARC, it makes no sense to have hash table of the same size (or even bigger on Linux). Reduce it to 1/8 of ARC's one, still leaving some slack, assuming higher I/O rate via dbuf cache than via ARC. Remove padding from ARC hash locks array. The idea behind padding is to avoid false sharing between locks. It would have sense if there would be a limited number of very busy locks. But since we have no limit on the number, using the same memory for more locks we can achieve even lower lock contention with the same false sharing, or we can use less memory for the same contention level. Dbuf hash locks never had this padding. Reduce number of hash locks from 8192 to 2048. The number is still big enough to not cause contention, but reduced memory size improves cache hit rate for mutex_tryenter() in ARC eviction thread, saving about 1% of the thread time. While there, move kmem_cache_free() out of dn_dbufs_mtx. Signed-off-by: Alexander Motin <mav@FreeBSD.org> Sponsored-By: iXsystems, Inc.
behlendorf
approved these changes
Jun 28, 2021
ahrens
approved these changes
Jun 29, 2021
behlendorf
pushed a commit
to behlendorf/zfs
that referenced
this pull request
Aug 23, 2021
With default dbuf cache size of 1/32 of ARC, it makes no sense to have hash table of the same size (or even bigger on Linux). Reduce it to 1/8 of ARC's one, still leaving some slack, assuming higher I/O rate via dbuf cache than via ARC. Remove padding from ARC hash locks array. The idea behind padding is to avoid false sharing between locks. It would have sense if there would be a limited number of very busy locks. But since we have no limit on the number, using the same memory for more locks we can achieve even lower lock contention with the same false sharing, or we can use less memory for the same contention level. Reduce number of hash locks from 8192 to 2048. The number is still big enough to not cause contention, but reduced memory size improves cache hit rate for mutex_tryenter() in ARC eviction thread, saving about 1% of the thread time. Reviewed-by: Matthew Ahrens <mahrens@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Alexander Motin <mav@FreeBSD.org> Closes openzfs#12289
behlendorf
pushed a commit
to behlendorf/zfs
that referenced
this pull request
Aug 24, 2021
With default dbuf cache size of 1/32 of ARC, it makes no sense to have hash table of the same size (or even bigger on Linux). Reduce it to 1/8 of ARC's one, still leaving some slack, assuming higher I/O rate via dbuf cache than via ARC. Remove padding from ARC hash locks array. The idea behind padding is to avoid false sharing between locks. It would have sense if there would be a limited number of very busy locks. But since we have no limit on the number, using the same memory for more locks we can achieve even lower lock contention with the same false sharing, or we can use less memory for the same contention level. Reduce number of hash locks from 8192 to 2048. The number is still big enough to not cause contention, but reduced memory size improves cache hit rate for mutex_tryenter() in ARC eviction thread, saving about 1% of the thread time. Reviewed-by: Matthew Ahrens <mahrens@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Alexander Motin <mav@FreeBSD.org> Closes openzfs#12289
behlendorf
pushed a commit
to behlendorf/zfs
that referenced
this pull request
Aug 24, 2021
With default dbuf cache size of 1/32 of ARC, it makes no sense to have hash table of the same size (or even bigger on Linux). Reduce it to 1/8 of ARC's one, still leaving some slack, assuming higher I/O rate via dbuf cache than via ARC. Remove padding from ARC hash locks array. The idea behind padding is to avoid false sharing between locks. It would have sense if there would be a limited number of very busy locks. But since we have no limit on the number, using the same memory for more locks we can achieve even lower lock contention with the same false sharing, or we can use less memory for the same contention level. Reduce number of hash locks from 8192 to 2048. The number is still big enough to not cause contention, but reduced memory size improves cache hit rate for mutex_tryenter() in ARC eviction thread, saving about 1% of the thread time. Reviewed-by: Matthew Ahrens <mahrens@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Alexander Motin <mav@FreeBSD.org> Closes openzfs#12289
behlendorf
pushed a commit
to behlendorf/zfs
that referenced
this pull request
Aug 24, 2021
With default dbuf cache size of 1/32 of ARC, it makes no sense to have hash table of the same size (or even bigger on Linux). Reduce it to 1/8 of ARC's one, still leaving some slack, assuming higher I/O rate via dbuf cache than via ARC. Remove padding from ARC hash locks array. The idea behind padding is to avoid false sharing between locks. It would have sense if there would be a limited number of very busy locks. But since we have no limit on the number, using the same memory for more locks we can achieve even lower lock contention with the same false sharing, or we can use less memory for the same contention level. Reduce number of hash locks from 8192 to 2048. The number is still big enough to not cause contention, but reduced memory size improves cache hit rate for mutex_tryenter() in ARC eviction thread, saving about 1% of the thread time. Reviewed-by: Matthew Ahrens <mahrens@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Alexander Motin <mav@FreeBSD.org> Closes openzfs#12289
behlendorf
pushed a commit
that referenced
this pull request
Aug 31, 2021
With default dbuf cache size of 1/32 of ARC, it makes no sense to have hash table of the same size (or even bigger on Linux). Reduce it to 1/8 of ARC's one, still leaving some slack, assuming higher I/O rate via dbuf cache than via ARC. Remove padding from ARC hash locks array. The idea behind padding is to avoid false sharing between locks. It would have sense if there would be a limited number of very busy locks. But since we have no limit on the number, using the same memory for more locks we can achieve even lower lock contention with the same false sharing, or we can use less memory for the same contention level. Reduce number of hash locks from 8192 to 2048. The number is still big enough to not cause contention, but reduced memory size improves cache hit rate for mutex_tryenter() in ARC eviction thread, saving about 1% of the thread time. Reviewed-by: Matthew Ahrens <mahrens@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Alexander Motin <mav@FreeBSD.org> Closes #12289
tonyhutter
pushed a commit
to tonyhutter/zfs
that referenced
this pull request
Sep 15, 2021
With default dbuf cache size of 1/32 of ARC, it makes no sense to have hash table of the same size (or even bigger on Linux). Reduce it to 1/8 of ARC's one, still leaving some slack, assuming higher I/O rate via dbuf cache than via ARC. Remove padding from ARC hash locks array. The idea behind padding is to avoid false sharing between locks. It would have sense if there would be a limited number of very busy locks. But since we have no limit on the number, using the same memory for more locks we can achieve even lower lock contention with the same false sharing, or we can use less memory for the same contention level. Reduce number of hash locks from 8192 to 2048. The number is still big enough to not cause contention, but reduced memory size improves cache hit rate for mutex_tryenter() in ARC eviction thread, saving about 1% of the thread time. Reviewed-by: Matthew Ahrens <mahrens@delphix.com> Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov> Signed-off-by: Alexander Motin <mav@FreeBSD.org> Closes openzfs#12289
13 tasks
ofaaland
added a commit
to LLNL/zfs
that referenced
this pull request
Mar 9, 2023
When running an HPSS repack (sending data from our old disk cache to our new disk cache) to our new ASP hardware HPSS admins and developers noticed a HUGE amount of CPU usage on our systems. They found that /proc/spl/kstat/zfs/dbufstats was reporting hash_chain_max of 13. HPSS uses zvols, whereas our Lustre systems do not. Possibly this is due to less uniform hashing of the blocks comprising zvols than non-zvol blocks, but this has not been verified. Increasing the size of the hash table dramatically improved performance, resulting in has_chain_max of 2. The hash table size was reduced recently, possibly too much. See openzfs#12289 An upstream patch will need to be written and landed, possibly to simply increase the dbuf cache size, or to dynamically size this hash, but this mitigation is sufficient for us in the meantime. Our hardware typically has plenty of RAM. Patch courtesy of Brian Behlendorf and Herb Wartens.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
With default dbuf cache size of 1/32 of ARC, it makes no sense to have
hash table of the same size (or even bigger on Linux). Reduce it to
1/8 of ARC's one, still leaving some slack, assuming higher I/O rate
via dbuf cache than via ARC.
Remove padding from ARC hash locks array. The idea behind padding
is to avoid false sharing between locks. It would have sense if
there would be a limited number of very busy locks. But since we
have no limit on the number, using the same memory for more locks we
can achieve even lower lock contention with the same false sharing,
or we can use less memory for the same contention level. Dbuf hash
locks never had this padding.
Reduce number of hash locks from 8192 to 2048. The number is still
big enough to not cause contention, but reduced memory size improves
cache hit rate for mutex_tryenter() in ARC eviction thread, saving
about 1% of the thread time.
While there, move kmem_cache_free() out of dn_dbufs_mtx.
How Has This Been Tested?
On 80-thread FreeBSD system with 768GB of RAM ZFS memory usage on boot reduced from 2GB to 1.1GB.
IOPS on heavy mostly uncached 4KB ZVOLs read test increased from 638K to 649K, while profiler shows reduction of CPU time spent in mutex_tryenter() from 15% to 14.2% of ARC reclaim thread.
Types of changes
Checklist:
Signed-off-by.