Performance regression in PS 5.5.19+ in flushing

Bug #1037343 reported by Vadim Tkachenko
14
This bug affects 3 people
Affects Status Importance Assigned to Milestone
Percona Server moved to https://jira.percona.com/projects/PS
Invalid
High
Unassigned

Bug Description

This bug is created to report observed problem and to keep track of ongoing research.

In Percona Server 5.5.19 and newer we have regression in flushing area.
In intensive tpcc-mysql benchmark I am getting long periods with 0 tps.

It is not the case in 5.5.18.

I suspect the problem comes from new buf_flush_try_neighbors function,
which introduces innodb_flush_neigbor=cont.
I suspect that innodb_flush_neigbor=0 got broken.

I am going to check it by replacing buf_flush_try_neighbors in 5.5.25 with the old one.

Tags: i25632
Changed in percona-server:
milestone: none → 5.5.28-29.0
Revision history for this message
Vadim Tkachenko (vadim-tk) wrote :

Unfortunately just using old version of buf_flush_try_neighbors does not help.
However it is very puzzling what would cause this regression, it requires an additional research.

Revision history for this message
Vadim Tkachenko (vadim-tk) wrote :

PS 5.5.27 also has this issue, just to check the latest version.

tags: added: i25632
Revision history for this message
Vadim Tkachenko (vadim-tk) wrote :

This problem is more complicated than I anticipate and actually consist of two part.

First part before Percona Server 5.5.25a (including) is that
 innodb_flush_neigbor=0 is not translated internally into 0 value, but rather into 3.
so
 innodb_flush_neigbor=0 is not equal innodb_flush_neigbor=none,

so the proper workaround is to use
 innodb_flush_neigbor=none.

However staring Percona Server 5.5.27, there an additional regression, which
I need to investigate further.

Revision history for this message
Laurynas Biveinis (laurynas-biveinis) wrote :

Handling the broken variables in bug 1039384.

Revision history for this message
Vadim Tkachenko (vadim-tk) wrote :

in fact with innodb_flush_neigbor=none
I do not see a regression in Percona Server 5.5.27 anymore.

I consider the problem resolved.

Stewart Smith (stewart)
no longer affects: percona-server/5.5
Changed in percona-server:
status: Confirmed → Invalid
Revision history for this message
Shahriyar Rzayev (rzayev-sehriyar) wrote :

Percona now uses JIRA for bug reports so this bug report is migrated to: https://jira.percona.com/browse/PS-577

To post a comment you must log in.
This report contains Public information  
Everyone can see this information.

Other bug subscribers

Remote bug watches

Bug watches keep track of this bug in other bug trackers.