Details
-
Improvement
-
Resolution: Unresolved
-
Major
-
6.0.4, 6.6.0
-
DOC-S15-May04
Description
During some recent investigation of the impact of compaction on front-end GET operations (MB-16750), we observed significant differences in performance between the ext4 and XFS filesystems.
Ask
We should summarise this to users - at the very least mention that XFS has been observed to offer significant performance improvements during compaction.
Background details (for future reference - not necessary to detail this in the docs)
In the test scenario, we had a workload with 50% / 50% GET/SET operations which had a cache miss rate of ~15-20%. Auto-compaction was enabled (at 30%), and in this environment compaction was essentially running constantly (spinning disks in a RAID array). I compared the results when the data partition was formatted as ext4 and XFS.
ext4
XFS
The drops in ext4 performance corresponded to when compaction was running. You can see that compaction has a marked impact (~14%) in ops per second when ext4 is used. When XFS is used, compaction has no negligible impact (essentially none is seen here).
Background Details
What appears to be happening is that ext4 has significantly larger bg_fetch times (time to fetch an item from the couchstore files on disk) compared to XFS. Note that in the configured environment no actual reads to the physical disk were occurring, as the machine had sufficient RAM to cache all disk blocks in RAM.