Project

General

Profile

Actions

Bug #59583

open

osd: Higher client latency observed with mclock 'high_client_ops' profile during recovery/backfills ops with EC backend (2+1) on SSDs

Added by Sridhar Seshasayee about 1 year ago. Updated about 1 year ago.

Status:
New
Priority:
Normal
Category:
Performance/Resource Usage
Target version:
-
% Done:

0%

Source:
Tags:
Backport:
Regression:
No
Severity:
3 - minor
Reviewed:
Affected Versions:
ceph-qa-suite:
Component(RADOS):
OSD
Pull request ID:
Crash signature (v1):
Crash signature (v2):

Description

Recovery/backfill testing was performed with OSDs on SSDs and with Erasure Coded backend. Tests with 'high_client_ops' mClock profile consistently showed higher client latency (avg & percentile) when compared to WPQ and other mClock profiles. The expectation is that with 'high_client_ops' profile, the client latency (avg and percentile) should be the lowest.

This bug is to track & investigate why the measured client latency with 'high_client_ops' profile is the highest.

The client latency comparison graph and the backfill rates graph across WPQ and mClock profiles are attached.


Files

Actions #1

Updated by Sridhar Seshasayee about 1 year ago

  • Subject changed from osd: Higher client latency observed with mclock 'high_client_ops' profile during recovery/backfills ops with EC backend (2+1) to osd: Higher client latency observed with mclock 'high_client_ops' profile during recovery/backfills ops with EC backend (2+1) on SSDs
Actions

Also available in: Atom PDF