FreeBSD 10.0-BETA1 #8 r256765M spend too much time in locks
Vitalij Satanivskij
satan at ukr.net
Thu Oct 24 07:48:36 UTC 2013
Hello.
After upgrading system from old current (r245701) to fresh current r255173 (than switch to stable/10 r256765M)
found some strange system behavior:
Diff from r256765M anf r256765 is
Index: sys/cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c
===================================================================
--- sys/cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c (revision 256765)
+++ sys/cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c (working copy)
@@ -5147,7 +5147,7 @@
len = l2hdr->b_asize;
cdata = zio_data_buf_alloc(len);
csize = zio_compress_data(ZIO_COMPRESS_LZ4, l2hdr->b_tmp_cdata,
- cdata, l2hdr->b_asize, (size_t)SPA_MINBLOCKSIZE);
+ cdata, l2hdr->b_asize, (size_t)(1ULL << l2hdr->b_dev->l2ad_vdev->vdev_ashift));
if (csize == 0) {
/* zero block, indicate that there's nothing to write */
But same situation was befor patch.
System load to high
CPU: 6.8% user, 0.0% nice, 57.3% system, 0.8% interrupt, 35.1% idle
hotkernel (dtrace script) says
kernel`__mtx_unlock_flags 292 0.3%
kernel`__mtx_lock_flags 315 0.3%
zfs.ko`lzjb_compress 349 0.3%
kernel`__rw_wlock_hard 686 0.7%
kernel`spinlock_exit 1050 1.0%
kernel`vmem_xalloc 7516 7.3%
kernel`_sx_xlock_hard 8664 8.5%
kernel`acpi_cpu_c1 9737 9.5%
kernel`cpu_idle 20189 19.7%
kernel`__mtx_lock_sleep 45952 44.9%
Trying to understand where is a problem I'm build kernel with lock profiling.
and get some data (for one minute )
(file attached)
using agregation find most lock's is in
14,159818 /usr/src/sys/kern/subr_vmem.c:1128(sleep mutex:kmem arena)
9,553523 /usr/src/sys/modules/zfs/../../cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c:1597(sx:buf_hash_table.ht_locks[i].ht_lock)
8,386943 /usr/src/sys/modules/zfs/../../cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c:3541(sx:l2arc_buflist_mtx)
8,110206 /usr/src/sys/kern/subr_vmem.c:1230(sleep mutex:kmem arena)
5,909429 /usr/src/sys/modules/zfs/../../cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c:1969(sx:arc_mru->arcs_locks[i].arcs_lock)
5,452206 /usr/src/sys/modules/zfs/../../cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c:1969(sx:arc_mfu->arcs_locks[i].arcs_lock)
5,050224 /usr/src/sys/modules/zfs/../../cddl/contrib/opensolaris/uts/common/fs/zfs/txg.c:303(sx:tx->tx_cpu[c].tc_open_lock)
4,232819 /usr/src/sys/modules/zfs/../../cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c:891(sx:buf_hash_table.ht_locks[i].ht_lock)
4,211348 /usr/src/sys/kern/vfs_subr.c:2101(lockmgr:zfs)
4,011656 /usr/src/sys/modules/zfs/../../cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c:862(sx:buf_hash_table.ht_locks[i].ht_lock)
3,823698 /usr/src/sys/modules/zfs/../../cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c:2009(sx:arc_eviction_mtx)
2,697344 /usr/src/sys/modules/zfs/../../cddl/contrib/opensolaris/uts/common/fs/zfs/dbuf.c:126(sx:h->hash_mutexes[i])
2,343242 /usr/src/sys/modules/zfs/../../cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c:1256(sx:arc_mfu->arcs_locks[i].arcs_lock)
1,752713 /usr/src/sys/kern/vfs_lookup.c:707(lockmgr:zfs)
1,580856 /usr/src/sys/modules/zfs/../../cddl/contrib/opensolaris/uts/common/fs/zfs/zfs_znode.c:1136(sx:zfsvfs->z_hold_mtx[i])
1,534242 /usr/src/sys/modules/zfs/../../cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c:1291(sx:arc_mfu->arcs_locks[i].arcs_lock)
1,331583 /usr/src/sys/modules/zfs/../../cddl/contrib/opensolaris/uts/common/fs/zfs/dbuf.c:129(sx:db->db_mtx)
1,105058 /usr/src/sys/modules/zfs/../../cddl/contrib/opensolaris/uts/common/fs/zfs/vdev_queue.c:427(sx:vq->vq_lock)
1,080855 /usr/src/sys/modules/zfs/../../cddl/contrib/opensolaris/uts/common/fs/zfs/vdev_queue.c:396(sx:vq->vq_lock)
0,858568 /usr/src/sys/kern/vfs_cache.c:488(rw:Name Cache)
0,831652 /usr/src/sys/vm/vm_kern.c:344(rw:kmem vm object)
0,815439 /usr/src/sys/modules/zfs/../../cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c:1845(sx:buf_hash_table.ht_locks[i].ht_lock)
0,812613 /usr/src/sys/modules/zfs/../../cddl/contrib/opensolaris/uts/common/fs/zfs/arc.c:1256(sx:arc_mru->arcs_locks[i].arcs_lock)
0,794274 /usr/src/sys/modules/zfs/../../cddl/contrib/opensolaris/uts/common/fs/zfs/zfs_vnops.c:1529(lockmgr:zfs)
0,669845 /usr/src/sys/vm/uma_core.c:2097(sleep mutex:zio_cache)
Short system description
CPU E5-1650
MEM 128Gb ddr3-1600
Storage subsystem
36x1Tb WD RE4 drives on LSI SAS2308 Controler
3x180Gb Intel ssd 530 series as l2 cache
POOL is 18 mirrors, two drives in ich mirror and 3 cache devices
eg.
....
mirror-14 ONLINE 0 0 0
gpt/disk28 ONLINE 0 0 0
gpt/disk29 ONLINE 0 0 0
mirror-15 ONLINE 0 0 0
gpt/disk30 ONLINE 0 0 0
gpt/disk31 ONLINE 0 0 0
mirror-16 ONLINE 0 0 0
gpt/disk32 ONLINE 0 0 0
gpt/disk33 ONLINE 0 0 0
mirror-17 ONLINE 0 0 0
gpt/disk34 ONLINE 0 0 0
gpt/disk35 ONLINE 0 0 0
cache
ada1 ONLINE 0 0 0
ada2 ONLINE 0 0 0
ada3 ONLINE 0 0 0
POOL have two ZFS
main with options (diffs from default) -
compression lz4
secondarycache all
sync disabled
Data size for it around 6Tb (compresed) eg
disk1 refcompressratio 1.56x -
disk1 written 5,99T -
disk1 logicalused 10,8T -
disk1 logicalreferenced 9,32T -
and another with options
recordsize 4K, before it was 32k, but internal software use data blocks mostly 4k so we try to change (without real data realocate )
compresion -s off
sync disabled
secondarycache all
DATA size on it around 1.4Tb
ARC configured to use max 80Gb
top most time looks like this
Mem: 14G Active, 13G Inact, 94G Wired, 497M Cache, 3297M Buf, 2214M Free
ARC: 80G Total, 2010M MFU, 70G MRU, 49M Anon, 3822M Header, 4288M Other
LA's - around 10-20 depend on day time.
zpool iostat disk1 1
capacity operations bandwidth
pool alloc free read write read write
---------- ----- ----- ----- ----- ----- -----
disk1 7,45T 8,86T 546 1,49K 16,4M 13,4M
disk1 7,45T 8,86T 272 3,91K 11,7M 27,4M
disk1 7,45T 8,86T 344 2,94K 7,26M 25,2M
disk1 7,45T 8,86T 224 2,02K 9,91M 21,8M
disk1 7,45T 8,86T 244 2,35K 8,23M 18,4M
disk1 7,45T 8,86T 245 2,54K 6,44M 23,4M
disk1 7,45T 8,86T 114 2,94K 3,28M 13,3M
disk1 7,45T 8,86T 288 4,43K 6,09M 33,5M
disk1 7,45T 8,86T 157 1,26K 2,98M 15,7M
disk1 7,45T 8,86T 94 842 3,07M 13,6M
disk1 7,45T 8,86T 327 1,71K 9,63M 8,21M
disk1 7,45T 8,86T 237 1,81K 5,73M 29,3M
disk1 7,45T 8,86T 247 3,47K 5,17M 29,6M
disk1 7,45T 8,86T 165 2,37K 3,22M 16,7M
disk1 7,45T 8,86T 155 3,23K 3,27M 23,9M
Strange as timeout seted up to 10sec's.
What intresting - after reboot system work fine for some time, at last while ARC not become 80G size.
Low limit for arc is 40gb, strange but old system can take memory from arc eg like this
Mem: 32G Active, 8797M Inact, 78G Wired, 2370M Cache, 209M Buf, 3977M Free
ARC: 43G Total, 2204M MFU, 28G MRU, 135M Anon, 7898M Header, 5301M Other
On new ARC getting to it max allowed size.
So for now question is, what it can be, what we can try to improve system perfomance and so on?
More information about the freebsd-hackers
mailing list