Actions
Bug #47344
closedosd: Poor client IO throughput/latency observed with dmclock scheduler during background recovery ops.
Status:
Resolved
Priority:
Normal
Assignee:
Category:
Performance/Resource Usage
Target version:
-
% Done:
0%
Source:
Tags:
Backport:
Regression:
No
Severity:
3 - minor
Reviewed:
Affected Versions:
ceph-qa-suite:
Component(RADOS):
Pull request ID:
Crash signature (v1):
Crash signature (v2):
Description
Regardless of the higher weightage given to client IO when compared to
recovery IO, poor client throughput/latency is observed.
Analysis revealed that recovery IO delivered via MOSDPGPush/MOSDPGPushReply
messages are always pushed to the "immediate" queue regardless of the
priority. This resulted in the starvation of the scheduler queue containing
client Ops and therefore the poor performance observed on the client.
The fix involves properly categorizing recovery ops and pushing them into
the scheduler queue so that the QoS requirements are met.
Actions