This patchset started out as a "convert the buffer cache to rbtrees"
patch, and just gew from there as I peeled the onion from one
bottleneck to another. The second version of this patch does not go
as far as the first version - it drops the more radical changes as
they are not ready for integration yet.
I dropped the RCU inode cache lookups because it is not well enough
tested yet. The lock contention reductions allowed by the RCU inode
cache lookups are replaced by more efficient lookup mechanisms
during inode cache walking - using batching mechanisms as originally
suggested by Nick Piggin. This code is more efficient than Nick's
proof of concept as it uses batched gang lookups on the radix trees.
These batched lookups show almost the same performance improvement
as the RCU lookup did but without changing the locking algorithms at
all. This batching is necessary for efficient reclaim walks
regardless of whether the sync walk is protected by RCU or the
current rwlock.
I also dropped the no-page-cache conversion patches for the buffer
cache as they need more work and testing before they are ready.
The shrinker rework improves parallel unlink performance
substantially more than just single threading the shrinker execution
and does not have the OOM problems that single threading the
shrinker had. It avoids the OOM problems by ensuring that every
shrinker call does some work or sleeps while waiting for an AG to do
some work on. The lookup optimisations done for gang lookups ensure
that the scanning is as efficient as possible, so overall shrinker
overhead has gone down significantly.
Performance numbers here are 8-way fs_mark create to 50M files, and
8-way rm -rf to remove the files created.
wall time fs_mark rate
2.6.36-rc4:
create: 13m10s 65k file/s
unlink: 23m58s N/A
2.6.36-rc4 + v1-patchset:
create: 9m47s 95k files/s
unlink: 14m16s N/A
2.6.36-rc3 + v2-patchset:
create: 10m32s 85k file/s
unlink: 11m49s N/A
So as you can see, the new patch set is a little slower on creates than
the previous version, but is still much faster than vanilla. The
unlink test is much faster than both vanilla and the previous version
thanks to the rework of the reclaim lookup and shrinker operations.
A breif description of the changes are:
xfs: reduce the number of CIL lock round trips during commit
- reduces lock traffic on the xc_cil_lock by two orders of
magnitude
xfs: remove debug assert for per-ag reference counting
xfs: lockless per-ag lookups
- hottest lock in the system with buffer cache rbtree path
- converted to use RCU.
xfs: don't use vfs writeback for pure metadata modifications
- inode writeback does not keep up with dirtying 100,000
inodes a second. Avoids the superblock dirty list where
possible by using the AIL as the age-order flusher.
xfs: rename xfs_buf_get_nodaddr to be more appropriate
xfs: introduced uncached buffer read primitve
xfs: store xfs_mount in the buftarg instead of in the xfs_buf
xfs: kill XBF_FS_MANAGED buffers
xfs: use unhashed buffers for size checks
xfs: remove buftarg hash for external devices
- preparatory buffer cache API cleanup patches
xfs: split inode AG walking into separate code for reclaim
xfs: implement batched inode lookups for AG walking
xfs: batch inode reclaim lookup
xfs: serialise inode reclaim within an AG
- inode cache shrinker rework
xfs: convert buffer cache hash to rbtree
xfs; pack xfs_buf structure more tightly
- conversion of buffer cache to use rbtrees for caches.
Version 2:
o dropped inode cache RCU/spinlock conversion (needs more testing)
o dropped buffer cache LRU/no page cache conversion (needs more testing)
o added CIL item insertion cleanup as suggested by Christoph.
o added flags to xfs_buf_get_uncached() and xfs_buf_read_uncached()
to control memory allocation flags.
o cleaned up buffer page allocation failure path
o reworked inode reclaim shrinker scalability
- separated reclaim AG walk from sync walks
- implemented batch lookups for both sync and reclaim walks
- added per-ag reclaim serialisation locks and traversal
cursors
The patches are available in the following git tree. The branch is
based on the current OSS xfs tree, and as such is based on
2.6.36-rc3 (which is why the tests were run on -rc3 instead of
-rc4).
Note: The branch also contains Christoph's btree + dquot cleanups;
they have been sitting in my tree being tested for quite some time
now and have been tested all through this metadata scalability work.
The following changes since commit 0e251465b06b75dfed16b9373c25cce85eeda484:
xfs: log IO completion workqueue is a high priority queue (2010-09-10 10:15:51 -0500)
are available in the git repository at:
git://git.kernel.org/pub/scm/linux/kernel/git/dgc/xfsdev.git metadata-scale
Christoph Hellwig (3):
xfs: remove the ->kill_root btree operation
xfs: simplify xfs_qm_dqusage_adjust
xfs: stop using xfs_qm_dqtobp in xfs_qm_dqflush
Dave Chinner (16):
xfs: reduce the number of CIL lock round trips during commit
xfs: remove debug assert for per-ag reference counting
xfs: lockless per-ag lookups
xfs: don't use vfs writeback for pure metadata modifications
xfs: rename xfs_buf_get_nodaddr to be more appropriate
xfs: introduced uncached buffer read primitve
xfs: store xfs_mount in the buftarg instead of in the xfs_buf
xfs: kill XBF_FS_MANAGED buffers
xfs: use unhashed buffers for size checks
xfs: remove buftarg hash for external devices
xfs: split inode AG walking into separate code for reclaim
xfs: implement batched inode lookups for AG walking
xfs: batch inode reclaim lookup
xfs: serialise inode reclaim within an AG
xfs: convert buffer cache hash to rbtree
xfs; pack xfs_buf structure more tightly
fs/xfs/linux-2.6/xfs_buf.c | 198 ++++++++++++----------
fs/xfs/linux-2.6/xfs_buf.h | 50 +++---
fs/xfs/linux-2.6/xfs_ioctl.c | 2 +-
fs/xfs/linux-2.6/xfs_iops.c | 55 ++++--
fs/xfs/linux-2.6/xfs_super.c | 8 +-
fs/xfs/linux-2.6/xfs_sync.c | 378 +++++++++++++++++++++++-----------------
fs/xfs/linux-2.6/xfs_sync.h | 5 +-
fs/xfs/linux-2.6/xfs_trace.h | 4 +-
fs/xfs/quota/xfs_dquot.c | 164 ++++++++----------
fs/xfs/quota/xfs_qm.c | 203 +++++++---------------
fs/xfs/quota/xfs_qm_syscalls.c | 29 ++--
fs/xfs/xfs_ag.h | 9 +
fs/xfs/xfs_alloc_btree.c | 33 ----
fs/xfs/xfs_attr.c | 36 ++--
fs/xfs/xfs_btree.c | 52 +++++-
fs/xfs/xfs_btree.h | 14 +--
fs/xfs/xfs_buf_item.c | 3 +-
fs/xfs/xfs_fsops.c | 11 +-
fs/xfs/xfs_ialloc_btree.c | 33 ----
fs/xfs/xfs_inode.h | 1 +
fs/xfs/xfs_inode_item.c | 9 -
fs/xfs/xfs_log.c | 3 +-
fs/xfs/xfs_log_cil.c | 232 ++++++++++++++-----------
fs/xfs/xfs_log_recover.c | 19 +-
fs/xfs/xfs_mount.c | 152 +++++++++--------
fs/xfs/xfs_mount.h | 2 +
fs/xfs/xfs_rename.c | 12 +-
fs/xfs/xfs_rtalloc.c | 29 ++--
fs/xfs/xfs_utils.c | 4 +-
fs/xfs/xfs_vnodeops.c | 12 +-
30 files changed, 884 insertions(+), 878 deletions(-)
_______________________________________________
xfs mailing list
xfs@xxxxxxxxxxx
http://oss.sgi.com/mailman/listinfo/xfs
