xfs: make inode reclaim almost non-blocking
Now that dirty inode writeback doesn't cause read-modify-write cycles on the inode cluster buffer under memory pressure, the need to throttle memory reclaim to the rate at which we can clean dirty inodes goes away. That is due to the fact that we no longer thrash inode cluster buffers under memory pressure to clean dirty inodes. This means inode writeback no longer stalls on memory allocation or read IO, and hence can be done asynchronously without generating memory pressure. As a result, blocking inode writeback in reclaim is no longer necessary to prevent reclaim priority windup as cleaning dirty inodes is no longer dependent on having memory reserves available for the filesystem to make progress reclaiming inodes. Hence we can convert inode reclaim to be non-blocking for shrinker callouts, both for direct reclaim and kswapd. On a vanilla kernel, running a 16-way fsmark create workload on a 4 node/16p/16GB RAM machine, I can reliably pin 14.75GB of RAM via userspace mlock(). The OOM killer gets invoked at 15GB of pinned RAM. Without the inode cluster pinning, this non-blocking reclaim patch triggers premature OOM killer invocation with the same memory pinning, sometimes with as much as 45% of RAM being free. It's trivially easy to trigger the OOM killer when reclaim does not block. With pinning inode clusters in RAM and then adding this patch, I can reliably pin 14.5GB of RAM and still have the fsmark workload run to completion. The OOM killer gets invoked 14.75GB of pinned RAM, which is only a small amount of memory less than the vanilla kernel. It is much more reliable than just with async reclaim alone. simoops shows that allocation stalls go away when async reclaim is used. Vanilla kernel: Run time: 1924 seconds Read latency (p50: 3,305,472) (p95: 3,723,264) (p99: 4,001,792) Write latency (p50: 184,064) (p95: 553,984) (p99: 807,936) Allocation latency (p50: 2,641,920) (p95: 3,911,680) (p99: 4,464,640) work rate = 13.45/sec (avg 13.44/sec) (p50: 13.46) (p95: 13.58) (p99: 13.70) alloc stall rate = 3.80/sec (avg: 2.59) (p50: 2.54) (p95: 2.96) (p99: 3.02) With inode cluster pinning and async reclaim: Run time: 1924 seconds Read latency (p50: 3,305,472) (p95: 3,715,072) (p99: 3,977,216) Write latency (p50: 187,648) (p95: 553,984) (p99: 789,504) Allocation latency (p50: 2,748,416) (p95: 3,919,872) (p99: 4,448,256) work rate = 13.28/sec (avg 13.32/sec) (p50: 13.26) (p95: 13.34) (p99: 13.34) alloc stall rate = 0.02/sec (avg: 0.02) (p50: 0.01) (p95: 0.03) (p99: 0.03) Latencies don't really change much, nor does the work rate. However, allocation almost never stalls with these changes, whilst the vanilla kernel is sometimes reporting 20 stalls/s over a 60s sample period. This difference is due to inode reclaim being largely non-blocking now. IOWs, once we have pinned inode cluster buffers, we can make inode reclaim non-blocking without a major risk of premature and/or spurious OOM killer invocation, and without any changes to memory reclaim infrastructure. Signed-off-by: Dave Chinner <dchinner@redhat.com> Reviewed-by: Amir Goldstein <amir73il@gmail.com> Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com> Reviewed-by: Brian Foster <bfoster@redhat.com> Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
This commit is contained in:
parent
298f7bec50
commit
993f951f50
|
@ -1402,7 +1402,7 @@ xfs_reclaim_inodes_nr(
|
|||
xfs_reclaim_work_queue(mp);
|
||||
xfs_ail_push_all(mp->m_ail);
|
||||
|
||||
return xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK | SYNC_WAIT, &nr_to_scan);
|
||||
return xfs_reclaim_inodes_ag(mp, SYNC_TRYLOCK, &nr_to_scan);
|
||||
}
|
||||
|
||||
/*
|
||||
|
|
Loading…
Reference in New Issue