Project

General

Profile

Actions

Bug #4206

closed

concurrent rados bench processes don't work well for seq reads

Added by Mark Nelson about 11 years ago. Updated over 9 years ago.

Status:
Resolved
Priority:
Normal
Assignee:
-
Category:
-
Target version:
-
% Done:

0%

Source:
Development
Tags:
Backport:
Regression:
Severity:
3 - minor
Reviewed:
Affected Versions:
ceph-qa-suite:
Pull request ID:
Crash signature (v1):
Crash signature (v2):

Description

Running concurrent rados bench processes results in various different problems when performing reads from the same pool. I've seen times where the same objects get almost at the same time resulting in reads from pagecache and other times where IO errors are reported and rados bench dies. I assume that there isn't a distinct object prefix for each rados bench instance's objects.

It would be nice to allow the user to assign one for both writes and reads so that concurrent rados bench instances could be run with their own distinct set of objects and not have to resort to hacks like having individual pools for every single rados bench instance being run. That's what I'm doing now, but it does not scale well once the total number of PGs exceeds a certain number (somewhere in the 80-100k region).

Actions #1

Updated by Sage Weil about 11 years ago

There is a distinct prefix.. can you include teh exact commands that (sometimes) reproduce this?

Actions #2

Updated by Greg Farnum about 11 years ago

I don't think the prefix is used when doing reads though — just the last benchmark run which finished is picked. We should add an interface for reading from a different set if you know the prefix.

I can't imagine why you'd be getting IO errors out of rados bench at any point, though! Next time you see that please keep some logs.

Actions #3

Updated by Loïc Dachary over 9 years ago

  • Status changed from New to Resolved

Implemented by 308758b7878c48ab64caf71ff646e057c2c1c5aa

Actions

Also available in: Atom PDF