Actions
Bug #24909
closedRBD client IOPS pool stats are incorrect (2x higher; includes IO hints as an op)
Status:
Resolved
Priority:
Normal
Assignee:
Jason Dillaman
Category:
-
Target version:
-
% Done:
0%
Source:
Tags:
Backport:
mimic,luminous
Regression:
No
Severity:
3 - minor
Reviewed:
Affected Versions:
ceph-qa-suite:
Component(RADOS):
Pull request ID:
Crash signature (v1):
Crash signature (v2):
Description
While running performance testing with Ceph metrics gathering statistics on the cluster, I noticed that while my RBD workload was generating ~7,000 writes/second to the cluster, Ceph metrics was showing my client is writing ~14,000 writes/second. Since these stats are just extracted from the PG stats, I looked into how those are calculated and it appears that "CEPH_OSD_OP_SETALLOCHINT" counts as a write op in terms of stat generation. Since RBD will send an alloc hint w/ each write op, it results in effectively double the write ops/second reported.
Actions