Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Optimize microzaps. #14039

Merged
merged 1 commit into from
Oct 20, 2022
Merged

Optimize microzaps. #14039

merged 1 commit into from
Oct 20, 2022

Conversation

amotin
Copy link
Member

@amotin amotin commented Oct 17, 2022

Microzap on-disk format does not include a hash tree, expecting one to be built in RAM during mzap_open(). The built tree is linked to DMU user buffer, freed when original DMU buffer is dropped from cache. I've found that workloads accessing many large directories and having active eviction from DMU cache spend significant amount of time building and then destroying the trees. I've also found that for each 64 byte mzap element additional 64 byte tree element is allocated, that is a waste of memory and CPU caches.

Improve memory efficiency of the hash tree by switching from AVL-tree to B-tree. It allows to save 24 bytes per element just on pointers. Save 32 bits on mze_hash by storing only upper 32 bits since lower 32 bits are always zero for microzaps. Save 16 bits on mze_chunkid, since microzap can never have so many elements. Respectively with the 16 bits there can be no more than 16 bits of collision differentiators. As result, struct mzap_ent now drops from 48 (rounded to 64) to 8 bytes.

Tune B-trees for small data. Reduce BTREE_CORE_ELEMS from 128 to 126 to allow struct zfs_btree_core in case of 8 byte elements to pack into 2KB instead of 4KB. Aside of the microzaps it should also help 32bit range trees. Allow custom B-tree leaf size to reduce memmove() time.

Split zap_name_alloc() into zap_name_alloc() and zap_name_init_str(). It allows to not waste time allocating/freeing memory when processing multiple names in a loop during mzap_open().

Together on a pool with 10K directories of 1800 files each and DMU cache limited to 128MB this reduces time of find . -name zzz by 41% from 7.63s to 4.47s, and saves additional ~30% of CPU time on cache reclamation.

Types of changes

  • Bug fix (non-breaking change which fixes an issue)
  • New feature (non-breaking change which adds functionality)
  • Performance enhancement (non-breaking change which improves efficiency)
  • Code cleanup (non-breaking change which makes code smaller or more readable)
  • Breaking change (fix or feature that would cause existing functionality to change)
  • Library ABI change (libzfs, libzfs_core, libnvpair, libuutil and libzfsbootenv)
  • Documentation (a change to man pages or other documentation)

Checklist:

avl_create(&zap->zap_m.zap_avl, mze_compare,
sizeof (mzap_ent_t), offsetof(mzap_ent_t, mze_node));
zfs_btree_create_custom(&zap->zap_m.zap_tree, mze_compare,
sizeof (mzap_ent_t), 512);
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could you add a comment explaining why 512?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Added. Hope it explains enough.

Microzap on-disk format does not include a hash tree, expecting one to
be built in RAM during mzap_open().  The built tree is linked to DMU
user buffer, freed when original DMU buffer is dropped from cache. I've
found that workloads accessing many large directories and having active
eviction from DMU cache spend significant amount of time building and
then destroying the trees.  I've also found that for each 64 byte mzap
element additional 64 byte tree element is allocated, that is a waste
of memory and CPU caches.

Improve memory efficiency of the hash tree by switching from AVL-tree
to B-tree.  It allows to save 24 bytes per element just on pointers.
Save 32 bits on mze_hash by storing only upper 32 bits since lower 32
bits are always zero for microzaps.  Save 16 bits on mze_chunkid, since
microzap can never have so many elements.  Respectively with the 16 bits
there can be no more than 16 bits of collision differentiators.  As
result, struct mzap_ent now drops from 48 (rounded to 64) to 8 bytes.

Tune B-trees for small data.  Reduce BTREE_CORE_ELEMS from 128 to 126
to allow struct zfs_btree_core in case of 8 byte elements to pack into
2KB instead of 4KB.  Aside of the microzaps it should also help 32bit
range trees.  Allow custom B-tree leaf size to reduce memmove() time.

Split zap_name_alloc() into zap_name_alloc() and zap_name_init_str().
It allows to not waste time allocating/freeing memory when processing
multiple names in a loop during mzap_open().

Together on a pool with 10K directories of 1800 files each and DMU
cache limited to 128MB this reduces time of `find . -name zzz` by 41%
from 7.63s to 4.47s, and saves additional ~30% of CPU time on the DMU
cache reclamation.

Signed-off-by:	Alexander Motin <[email protected]>
Sponsored by:	iXsystems, Inc.
@behlendorf behlendorf added Status: Accepted Ready to integrate (reviewed, tested) and removed Status: Code Review Needed Ready for review and testing labels Oct 19, 2022
@behlendorf behlendorf merged commit 9dcdee7 into openzfs:master Oct 20, 2022
andrewc12 pushed a commit to andrewc12/openzfs that referenced this pull request Oct 21, 2022
Microzap on-disk format does not include a hash tree, expecting one to
be built in RAM during mzap_open().  The built tree is linked to DMU
user buffer, freed when original DMU buffer is dropped from cache. I've
found that workloads accessing many large directories and having active
eviction from DMU cache spend significant amount of time building and
then destroying the trees.  I've also found that for each 64 byte mzap
element additional 64 byte tree element is allocated, that is a waste
of memory and CPU caches.

Improve memory efficiency of the hash tree by switching from AVL-tree
to B-tree.  It allows to save 24 bytes per element just on pointers.
Save 32 bits on mze_hash by storing only upper 32 bits since lower 32
bits are always zero for microzaps.  Save 16 bits on mze_chunkid, since
microzap can never have so many elements.  Respectively with the 16 bits
there can be no more than 16 bits of collision differentiators.  As
result, struct mzap_ent now drops from 48 (rounded to 64) to 8 bytes.

Tune B-trees for small data.  Reduce BTREE_CORE_ELEMS from 128 to 126
to allow struct zfs_btree_core in case of 8 byte elements to pack into
2KB instead of 4KB.  Aside of the microzaps it should also help 32bit
range trees.  Allow custom B-tree leaf size to reduce memmove() time.

Split zap_name_alloc() into zap_name_alloc() and zap_name_init_str().
It allows to not waste time allocating/freeing memory when processing
multiple names in a loop during mzap_open().

Together on a pool with 10K directories of 1800 files each and DMU
cache limited to 128MB this reduces time of `find . -name zzz` by 41%
from 7.63s to 4.47s, and saves additional ~30% of CPU time on the DMU
cache reclamation.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Matthew Ahrens <[email protected]>
Reviewed-by: Ryan Moeller <[email protected]>
Signed-off-by:	Alexander Motin <[email protected]>
Sponsored by:	iXsystems, Inc.
Closes openzfs#14039
andrewc12 pushed a commit to andrewc12/openzfs that referenced this pull request Oct 21, 2022
Microzap on-disk format does not include a hash tree, expecting one to
be built in RAM during mzap_open().  The built tree is linked to DMU
user buffer, freed when original DMU buffer is dropped from cache. I've
found that workloads accessing many large directories and having active
eviction from DMU cache spend significant amount of time building and
then destroying the trees.  I've also found that for each 64 byte mzap
element additional 64 byte tree element is allocated, that is a waste
of memory and CPU caches.

Improve memory efficiency of the hash tree by switching from AVL-tree
to B-tree.  It allows to save 24 bytes per element just on pointers.
Save 32 bits on mze_hash by storing only upper 32 bits since lower 32
bits are always zero for microzaps.  Save 16 bits on mze_chunkid, since
microzap can never have so many elements.  Respectively with the 16 bits
there can be no more than 16 bits of collision differentiators.  As
result, struct mzap_ent now drops from 48 (rounded to 64) to 8 bytes.

Tune B-trees for small data.  Reduce BTREE_CORE_ELEMS from 128 to 126
to allow struct zfs_btree_core in case of 8 byte elements to pack into
2KB instead of 4KB.  Aside of the microzaps it should also help 32bit
range trees.  Allow custom B-tree leaf size to reduce memmove() time.

Split zap_name_alloc() into zap_name_alloc() and zap_name_init_str().
It allows to not waste time allocating/freeing memory when processing
multiple names in a loop during mzap_open().

Together on a pool with 10K directories of 1800 files each and DMU
cache limited to 128MB this reduces time of `find . -name zzz` by 41%
from 7.63s to 4.47s, and saves additional ~30% of CPU time on the DMU
cache reclamation.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Matthew Ahrens <[email protected]>
Reviewed-by: Ryan Moeller <[email protected]>
Signed-off-by:	Alexander Motin <[email protected]>
Sponsored by:	iXsystems, Inc.
Closes openzfs#14039
andrewc12 pushed a commit to andrewc12/openzfs that referenced this pull request Oct 21, 2022
Microzap on-disk format does not include a hash tree, expecting one to
be built in RAM during mzap_open().  The built tree is linked to DMU
user buffer, freed when original DMU buffer is dropped from cache. I've
found that workloads accessing many large directories and having active
eviction from DMU cache spend significant amount of time building and
then destroying the trees.  I've also found that for each 64 byte mzap
element additional 64 byte tree element is allocated, that is a waste
of memory and CPU caches.

Improve memory efficiency of the hash tree by switching from AVL-tree
to B-tree.  It allows to save 24 bytes per element just on pointers.
Save 32 bits on mze_hash by storing only upper 32 bits since lower 32
bits are always zero for microzaps.  Save 16 bits on mze_chunkid, since
microzap can never have so many elements.  Respectively with the 16 bits
there can be no more than 16 bits of collision differentiators.  As
result, struct mzap_ent now drops from 48 (rounded to 64) to 8 bytes.

Tune B-trees for small data.  Reduce BTREE_CORE_ELEMS from 128 to 126
to allow struct zfs_btree_core in case of 8 byte elements to pack into
2KB instead of 4KB.  Aside of the microzaps it should also help 32bit
range trees.  Allow custom B-tree leaf size to reduce memmove() time.

Split zap_name_alloc() into zap_name_alloc() and zap_name_init_str().
It allows to not waste time allocating/freeing memory when processing
multiple names in a loop during mzap_open().

Together on a pool with 10K directories of 1800 files each and DMU
cache limited to 128MB this reduces time of `find . -name zzz` by 41%
from 7.63s to 4.47s, and saves additional ~30% of CPU time on the DMU
cache reclamation.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Matthew Ahrens <[email protected]>
Reviewed-by: Ryan Moeller <[email protected]>
Signed-off-by:	Alexander Motin <[email protected]>
Sponsored by:	iXsystems, Inc.
Closes openzfs#14039
andrewc12 pushed a commit to andrewc12/openzfs that referenced this pull request Oct 21, 2022
Microzap on-disk format does not include a hash tree, expecting one to
be built in RAM during mzap_open().  The built tree is linked to DMU
user buffer, freed when original DMU buffer is dropped from cache. I've
found that workloads accessing many large directories and having active
eviction from DMU cache spend significant amount of time building and
then destroying the trees.  I've also found that for each 64 byte mzap
element additional 64 byte tree element is allocated, that is a waste
of memory and CPU caches.

Improve memory efficiency of the hash tree by switching from AVL-tree
to B-tree.  It allows to save 24 bytes per element just on pointers.
Save 32 bits on mze_hash by storing only upper 32 bits since lower 32
bits are always zero for microzaps.  Save 16 bits on mze_chunkid, since
microzap can never have so many elements.  Respectively with the 16 bits
there can be no more than 16 bits of collision differentiators.  As
result, struct mzap_ent now drops from 48 (rounded to 64) to 8 bytes.

Tune B-trees for small data.  Reduce BTREE_CORE_ELEMS from 128 to 126
to allow struct zfs_btree_core in case of 8 byte elements to pack into
2KB instead of 4KB.  Aside of the microzaps it should also help 32bit
range trees.  Allow custom B-tree leaf size to reduce memmove() time.

Split zap_name_alloc() into zap_name_alloc() and zap_name_init_str().
It allows to not waste time allocating/freeing memory when processing
multiple names in a loop during mzap_open().

Together on a pool with 10K directories of 1800 files each and DMU
cache limited to 128MB this reduces time of `find . -name zzz` by 41%
from 7.63s to 4.47s, and saves additional ~30% of CPU time on the DMU
cache reclamation.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Matthew Ahrens <[email protected]>
Reviewed-by: Ryan Moeller <[email protected]>
Signed-off-by:	Alexander Motin <[email protected]>
Sponsored by:	iXsystems, Inc.
Closes openzfs#14039
amotin added a commit to truenas/zfs that referenced this pull request Oct 21, 2022
Microzap on-disk format does not include a hash tree, expecting one to
be built in RAM during mzap_open().  The built tree is linked to DMU
user buffer, freed when original DMU buffer is dropped from cache. I've
found that workloads accessing many large directories and having active
eviction from DMU cache spend significant amount of time building and
then destroying the trees.  I've also found that for each 64 byte mzap
element additional 64 byte tree element is allocated, that is a waste
of memory and CPU caches.

Improve memory efficiency of the hash tree by switching from AVL-tree
to B-tree.  It allows to save 24 bytes per element just on pointers.
Save 32 bits on mze_hash by storing only upper 32 bits since lower 32
bits are always zero for microzaps.  Save 16 bits on mze_chunkid, since
microzap can never have so many elements.  Respectively with the 16 bits
there can be no more than 16 bits of collision differentiators.  As
result, struct mzap_ent now drops from 48 (rounded to 64) to 8 bytes.

Tune B-trees for small data.  Reduce BTREE_CORE_ELEMS from 128 to 126
to allow struct zfs_btree_core in case of 8 byte elements to pack into
2KB instead of 4KB.  Aside of the microzaps it should also help 32bit
range trees.  Allow custom B-tree leaf size to reduce memmove() time.

Split zap_name_alloc() into zap_name_alloc() and zap_name_init_str().
It allows to not waste time allocating/freeing memory when processing
multiple names in a loop during mzap_open().

Together on a pool with 10K directories of 1800 files each and DMU
cache limited to 128MB this reduces time of `find . -name zzz` by 41%
from 7.63s to 4.47s, and saves additional ~30% of CPU time on the DMU
cache reclamation.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Matthew Ahrens <[email protected]>
Reviewed-by: Ryan Moeller <[email protected]>
Signed-off-by:	Alexander Motin <[email protected]>
Sponsored by:	iXsystems, Inc.
Closes openzfs#14039

(cherry picked from commit 9dcdee7)
snajpa pushed a commit to vpsfreecz/zfs that referenced this pull request Oct 22, 2022
Microzap on-disk format does not include a hash tree, expecting one to
be built in RAM during mzap_open().  The built tree is linked to DMU
user buffer, freed when original DMU buffer is dropped from cache. I've
found that workloads accessing many large directories and having active
eviction from DMU cache spend significant amount of time building and
then destroying the trees.  I've also found that for each 64 byte mzap
element additional 64 byte tree element is allocated, that is a waste
of memory and CPU caches.

Improve memory efficiency of the hash tree by switching from AVL-tree
to B-tree.  It allows to save 24 bytes per element just on pointers.
Save 32 bits on mze_hash by storing only upper 32 bits since lower 32
bits are always zero for microzaps.  Save 16 bits on mze_chunkid, since
microzap can never have so many elements.  Respectively with the 16 bits
there can be no more than 16 bits of collision differentiators.  As
result, struct mzap_ent now drops from 48 (rounded to 64) to 8 bytes.

Tune B-trees for small data.  Reduce BTREE_CORE_ELEMS from 128 to 126
to allow struct zfs_btree_core in case of 8 byte elements to pack into
2KB instead of 4KB.  Aside of the microzaps it should also help 32bit
range trees.  Allow custom B-tree leaf size to reduce memmove() time.

Split zap_name_alloc() into zap_name_alloc() and zap_name_init_str().
It allows to not waste time allocating/freeing memory when processing
multiple names in a loop during mzap_open().

Together on a pool with 10K directories of 1800 files each and DMU
cache limited to 128MB this reduces time of `find . -name zzz` by 41%
from 7.63s to 4.47s, and saves additional ~30% of CPU time on the DMU
cache reclamation.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Matthew Ahrens <[email protected]>
Reviewed-by: Ryan Moeller <[email protected]>
Signed-off-by:	Alexander Motin <[email protected]>
Sponsored by:	iXsystems, Inc.
Closes openzfs#14039
snajpa pushed a commit to vpsfreecz/zfs that referenced this pull request Oct 22, 2022
Microzap on-disk format does not include a hash tree, expecting one to
be built in RAM during mzap_open().  The built tree is linked to DMU
user buffer, freed when original DMU buffer is dropped from cache. I've
found that workloads accessing many large directories and having active
eviction from DMU cache spend significant amount of time building and
then destroying the trees.  I've also found that for each 64 byte mzap
element additional 64 byte tree element is allocated, that is a waste
of memory and CPU caches.

Improve memory efficiency of the hash tree by switching from AVL-tree
to B-tree.  It allows to save 24 bytes per element just on pointers.
Save 32 bits on mze_hash by storing only upper 32 bits since lower 32
bits are always zero for microzaps.  Save 16 bits on mze_chunkid, since
microzap can never have so many elements.  Respectively with the 16 bits
there can be no more than 16 bits of collision differentiators.  As
result, struct mzap_ent now drops from 48 (rounded to 64) to 8 bytes.

Tune B-trees for small data.  Reduce BTREE_CORE_ELEMS from 128 to 126
to allow struct zfs_btree_core in case of 8 byte elements to pack into
2KB instead of 4KB.  Aside of the microzaps it should also help 32bit
range trees.  Allow custom B-tree leaf size to reduce memmove() time.

Split zap_name_alloc() into zap_name_alloc() and zap_name_init_str().
It allows to not waste time allocating/freeing memory when processing
multiple names in a loop during mzap_open().

Together on a pool with 10K directories of 1800 files each and DMU
cache limited to 128MB this reduces time of `find . -name zzz` by 41%
from 7.63s to 4.47s, and saves additional ~30% of CPU time on the DMU
cache reclamation.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Matthew Ahrens <[email protected]>
Reviewed-by: Ryan Moeller <[email protected]>
Signed-off-by:	Alexander Motin <[email protected]>
Sponsored by:	iXsystems, Inc.
Closes openzfs#14039
snajpa pushed a commit to vpsfreecz/zfs that referenced this pull request Oct 23, 2022
Microzap on-disk format does not include a hash tree, expecting one to
be built in RAM during mzap_open().  The built tree is linked to DMU
user buffer, freed when original DMU buffer is dropped from cache. I've
found that workloads accessing many large directories and having active
eviction from DMU cache spend significant amount of time building and
then destroying the trees.  I've also found that for each 64 byte mzap
element additional 64 byte tree element is allocated, that is a waste
of memory and CPU caches.

Improve memory efficiency of the hash tree by switching from AVL-tree
to B-tree.  It allows to save 24 bytes per element just on pointers.
Save 32 bits on mze_hash by storing only upper 32 bits since lower 32
bits are always zero for microzaps.  Save 16 bits on mze_chunkid, since
microzap can never have so many elements.  Respectively with the 16 bits
there can be no more than 16 bits of collision differentiators.  As
result, struct mzap_ent now drops from 48 (rounded to 64) to 8 bytes.

Tune B-trees for small data.  Reduce BTREE_CORE_ELEMS from 128 to 126
to allow struct zfs_btree_core in case of 8 byte elements to pack into
2KB instead of 4KB.  Aside of the microzaps it should also help 32bit
range trees.  Allow custom B-tree leaf size to reduce memmove() time.

Split zap_name_alloc() into zap_name_alloc() and zap_name_init_str().
It allows to not waste time allocating/freeing memory when processing
multiple names in a loop during mzap_open().

Together on a pool with 10K directories of 1800 files each and DMU
cache limited to 128MB this reduces time of `find . -name zzz` by 41%
from 7.63s to 4.47s, and saves additional ~30% of CPU time on the DMU
cache reclamation.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Matthew Ahrens <[email protected]>
Reviewed-by: Ryan Moeller <[email protected]>
Signed-off-by:	Alexander Motin <[email protected]>
Sponsored by:	iXsystems, Inc.
Closes openzfs#14039
andrewc12 pushed a commit to andrewc12/openzfs that referenced this pull request Nov 9, 2022
Microzap on-disk format does not include a hash tree, expecting one to
be built in RAM during mzap_open().  The built tree is linked to DMU
user buffer, freed when original DMU buffer is dropped from cache. I've
found that workloads accessing many large directories and having active
eviction from DMU cache spend significant amount of time building and
then destroying the trees.  I've also found that for each 64 byte mzap
element additional 64 byte tree element is allocated, that is a waste
of memory and CPU caches.

Improve memory efficiency of the hash tree by switching from AVL-tree
to B-tree.  It allows to save 24 bytes per element just on pointers.
Save 32 bits on mze_hash by storing only upper 32 bits since lower 32
bits are always zero for microzaps.  Save 16 bits on mze_chunkid, since
microzap can never have so many elements.  Respectively with the 16 bits
there can be no more than 16 bits of collision differentiators.  As
result, struct mzap_ent now drops from 48 (rounded to 64) to 8 bytes.

Tune B-trees for small data.  Reduce BTREE_CORE_ELEMS from 128 to 126
to allow struct zfs_btree_core in case of 8 byte elements to pack into
2KB instead of 4KB.  Aside of the microzaps it should also help 32bit
range trees.  Allow custom B-tree leaf size to reduce memmove() time.

Split zap_name_alloc() into zap_name_alloc() and zap_name_init_str().
It allows to not waste time allocating/freeing memory when processing
multiple names in a loop during mzap_open().

Together on a pool with 10K directories of 1800 files each and DMU
cache limited to 128MB this reduces time of `find . -name zzz` by 41%
from 7.63s to 4.47s, and saves additional ~30% of CPU time on the DMU
cache reclamation.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Matthew Ahrens <[email protected]>
Reviewed-by: Ryan Moeller <[email protected]>
Signed-off-by:	Alexander Motin <[email protected]>
Sponsored by:	iXsystems, Inc.
Closes openzfs#14039
andrewc12 pushed a commit to andrewc12/openzfs that referenced this pull request Nov 9, 2022
Microzap on-disk format does not include a hash tree, expecting one to
be built in RAM during mzap_open().  The built tree is linked to DMU
user buffer, freed when original DMU buffer is dropped from cache. I've
found that workloads accessing many large directories and having active
eviction from DMU cache spend significant amount of time building and
then destroying the trees.  I've also found that for each 64 byte mzap
element additional 64 byte tree element is allocated, that is a waste
of memory and CPU caches.

Improve memory efficiency of the hash tree by switching from AVL-tree
to B-tree.  It allows to save 24 bytes per element just on pointers.
Save 32 bits on mze_hash by storing only upper 32 bits since lower 32
bits are always zero for microzaps.  Save 16 bits on mze_chunkid, since
microzap can never have so many elements.  Respectively with the 16 bits
there can be no more than 16 bits of collision differentiators.  As
result, struct mzap_ent now drops from 48 (rounded to 64) to 8 bytes.

Tune B-trees for small data.  Reduce BTREE_CORE_ELEMS from 128 to 126
to allow struct zfs_btree_core in case of 8 byte elements to pack into
2KB instead of 4KB.  Aside of the microzaps it should also help 32bit
range trees.  Allow custom B-tree leaf size to reduce memmove() time.

Split zap_name_alloc() into zap_name_alloc() and zap_name_init_str().
It allows to not waste time allocating/freeing memory when processing
multiple names in a loop during mzap_open().

Together on a pool with 10K directories of 1800 files each and DMU
cache limited to 128MB this reduces time of `find . -name zzz` by 41%
from 7.63s to 4.47s, and saves additional ~30% of CPU time on the DMU
cache reclamation.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Matthew Ahrens <[email protected]>
Reviewed-by: Ryan Moeller <[email protected]>
Signed-off-by:	Alexander Motin <[email protected]>
Sponsored by:	iXsystems, Inc.
Closes openzfs#14039
andrewc12 pushed a commit to andrewc12/openzfs that referenced this pull request Nov 9, 2022
Microzap on-disk format does not include a hash tree, expecting one to
be built in RAM during mzap_open().  The built tree is linked to DMU
user buffer, freed when original DMU buffer is dropped from cache. I've
found that workloads accessing many large directories and having active
eviction from DMU cache spend significant amount of time building and
then destroying the trees.  I've also found that for each 64 byte mzap
element additional 64 byte tree element is allocated, that is a waste
of memory and CPU caches.

Improve memory efficiency of the hash tree by switching from AVL-tree
to B-tree.  It allows to save 24 bytes per element just on pointers.
Save 32 bits on mze_hash by storing only upper 32 bits since lower 32
bits are always zero for microzaps.  Save 16 bits on mze_chunkid, since
microzap can never have so many elements.  Respectively with the 16 bits
there can be no more than 16 bits of collision differentiators.  As
result, struct mzap_ent now drops from 48 (rounded to 64) to 8 bytes.

Tune B-trees for small data.  Reduce BTREE_CORE_ELEMS from 128 to 126
to allow struct zfs_btree_core in case of 8 byte elements to pack into
2KB instead of 4KB.  Aside of the microzaps it should also help 32bit
range trees.  Allow custom B-tree leaf size to reduce memmove() time.

Split zap_name_alloc() into zap_name_alloc() and zap_name_init_str().
It allows to not waste time allocating/freeing memory when processing
multiple names in a loop during mzap_open().

Together on a pool with 10K directories of 1800 files each and DMU
cache limited to 128MB this reduces time of `find . -name zzz` by 41%
from 7.63s to 4.47s, and saves additional ~30% of CPU time on the DMU
cache reclamation.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Matthew Ahrens <[email protected]>
Reviewed-by: Ryan Moeller <[email protected]>
Signed-off-by:	Alexander Motin <[email protected]>
Sponsored by:	iXsystems, Inc.
Closes openzfs#14039
andrewc12 pushed a commit to andrewc12/openzfs that referenced this pull request Nov 9, 2022
Microzap on-disk format does not include a hash tree, expecting one to
be built in RAM during mzap_open().  The built tree is linked to DMU
user buffer, freed when original DMU buffer is dropped from cache. I've
found that workloads accessing many large directories and having active
eviction from DMU cache spend significant amount of time building and
then destroying the trees.  I've also found that for each 64 byte mzap
element additional 64 byte tree element is allocated, that is a waste
of memory and CPU caches.

Improve memory efficiency of the hash tree by switching from AVL-tree
to B-tree.  It allows to save 24 bytes per element just on pointers.
Save 32 bits on mze_hash by storing only upper 32 bits since lower 32
bits are always zero for microzaps.  Save 16 bits on mze_chunkid, since
microzap can never have so many elements.  Respectively with the 16 bits
there can be no more than 16 bits of collision differentiators.  As
result, struct mzap_ent now drops from 48 (rounded to 64) to 8 bytes.

Tune B-trees for small data.  Reduce BTREE_CORE_ELEMS from 128 to 126
to allow struct zfs_btree_core in case of 8 byte elements to pack into
2KB instead of 4KB.  Aside of the microzaps it should also help 32bit
range trees.  Allow custom B-tree leaf size to reduce memmove() time.

Split zap_name_alloc() into zap_name_alloc() and zap_name_init_str().
It allows to not waste time allocating/freeing memory when processing
multiple names in a loop during mzap_open().

Together on a pool with 10K directories of 1800 files each and DMU
cache limited to 128MB this reduces time of `find . -name zzz` by 41%
from 7.63s to 4.47s, and saves additional ~30% of CPU time on the DMU
cache reclamation.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Matthew Ahrens <[email protected]>
Reviewed-by: Ryan Moeller <[email protected]>
Signed-off-by:	Alexander Motin <[email protected]>
Sponsored by:	iXsystems, Inc.
Closes openzfs#14039
amotin added a commit to amotin/zfs that referenced this pull request Mar 2, 2023
Microzap on-disk format does not include a hash tree, expecting one to
be built in RAM during mzap_open().  The built tree is linked to DMU
user buffer, freed when original DMU buffer is dropped from cache. I've
found that workloads accessing many large directories and having active
eviction from DMU cache spend significant amount of time building and
then destroying the trees.  I've also found that for each 64 byte mzap
element additional 64 byte tree element is allocated, that is a waste
of memory and CPU caches.

Improve memory efficiency of the hash tree by switching from AVL-tree
to B-tree.  It allows to save 24 bytes per element just on pointers.
Save 32 bits on mze_hash by storing only upper 32 bits since lower 32
bits are always zero for microzaps.  Save 16 bits on mze_chunkid, since
microzap can never have so many elements.  Respectively with the 16 bits
there can be no more than 16 bits of collision differentiators.  As
result, struct mzap_ent now drops from 48 (rounded to 64) to 8 bytes.

Tune B-trees for small data.  Reduce BTREE_CORE_ELEMS from 128 to 126
to allow struct zfs_btree_core in case of 8 byte elements to pack into
2KB instead of 4KB.  Aside of the microzaps it should also help 32bit
range trees.  Allow custom B-tree leaf size to reduce memmove() time.

Split zap_name_alloc() into zap_name_alloc() and zap_name_init_str().
It allows to not waste time allocating/freeing memory when processing
multiple names in a loop during mzap_open().

Together on a pool with 10K directories of 1800 files each and DMU
cache limited to 128MB this reduces time of `find . -name zzz` by 41%
from 7.63s to 4.47s, and saves additional ~30% of CPU time on the DMU
cache reclamation.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Matthew Ahrens <[email protected]>
Reviewed-by: Ryan Moeller <[email protected]>
Signed-off-by:	Alexander Motin <[email protected]>
Sponsored by:	iXsystems, Inc.
Closes openzfs#14039

(cherry picked from commit 9dcdee7)
behlendorf pushed a commit that referenced this pull request Mar 2, 2023
Microzap on-disk format does not include a hash tree, expecting one to
be built in RAM during mzap_open().  The built tree is linked to DMU
user buffer, freed when original DMU buffer is dropped from cache. I've
found that workloads accessing many large directories and having active
eviction from DMU cache spend significant amount of time building and
then destroying the trees.  I've also found that for each 64 byte mzap
element additional 64 byte tree element is allocated, that is a waste
of memory and CPU caches.

Improve memory efficiency of the hash tree by switching from AVL-tree
to B-tree.  It allows to save 24 bytes per element just on pointers.
Save 32 bits on mze_hash by storing only upper 32 bits since lower 32
bits are always zero for microzaps.  Save 16 bits on mze_chunkid, since
microzap can never have so many elements.  Respectively with the 16 bits
there can be no more than 16 bits of collision differentiators.  As
result, struct mzap_ent now drops from 48 (rounded to 64) to 8 bytes.

Tune B-trees for small data.  Reduce BTREE_CORE_ELEMS from 128 to 126
to allow struct zfs_btree_core in case of 8 byte elements to pack into
2KB instead of 4KB.  Aside of the microzaps it should also help 32bit
range trees.  Allow custom B-tree leaf size to reduce memmove() time.

Split zap_name_alloc() into zap_name_alloc() and zap_name_init_str().
It allows to not waste time allocating/freeing memory when processing
multiple names in a loop during mzap_open().

Together on a pool with 10K directories of 1800 files each and DMU
cache limited to 128MB this reduces time of `find . -name zzz` by 41%
from 7.63s to 4.47s, and saves additional ~30% of CPU time on the DMU
cache reclamation.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Matthew Ahrens <[email protected]>
Reviewed-by: Ryan Moeller <[email protected]>
Signed-off-by:	Alexander Motin <[email protected]>
Sponsored by:	iXsystems, Inc.
Closes #14039

(cherry picked from commit 9dcdee7)
@amotin amotin deleted the mzap branch October 9, 2023 13:22
behlendorf added a commit that referenced this pull request Oct 13, 2023
New Features
- Block cloning (#13392)
- Linux container support (#14070, #14097, #12263)
- Scrub error log (#12812, #12355)
- BLAKE3 checksums (#12918)
- Corrective "zfs receive"
- Vdev and zpool user properties

Performance
- Fully adaptive ARC (#14359)
- SHA2 checksums (#13741)
- Edon-R checksums (#13618)
- Zstd early abort (#13244)
- Prefetch improvements (#14603, #14516, #14402, #14243, #13452)
- General optimization (#14121, #14123, #14039, #13680, #13613,
  #13606, #13576, #13553, #12789, #14925, #14948)

Signed-off-by: Brian Behlendorf <[email protected]>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
Status: Accepted Ready to integrate (reviewed, tested)
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants