Page MenuHomePhabricator

Cirrussearch: Enable performance governor on master-eligibles
Closed, ResolvedPublic

Description

While investigating cluster quorum issues in parent ticket, we've noticed that several master-eligibles, including all hosts in EQIAD, don't have the performance governor enabled:

powersave
[bking@cumin2003] ~$ sudo cumin A:cirrussearch-codfw-masters 'cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor | uniq'
15 hosts will be targeted:
cirrussearch[2061,2073,2076,2080-2081,2083-2084,2086,2092-2093,2100,2106,2108,2111,2115].codfw.wmnet
===== NODE GROUP =====
(11) cirrussearch[2061,2073,2076,2080-2081,2084,2086,2092-2093,2100,2111].codfw.wmnet
----- OUTPUT for command #1: 'cat /sys/devices..._governor | uniq' -----
powersave

[bking@cumin2003] ~$ sudo cumin A:cirrussearch-eqiad-masters 'cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor | uniq'
15 hosts will be targeted:
cirrussearch[1068,1073-1076,1081,1093-1095,1098,1100,1102,1118-1119,1122].eqiad.wmnet
----- OUTPUT for command #1: 'cat /sys/devices..._governor | uniq' -----
powersave
===== NODE GROUP =====
(3) cirrussearch[1118-1119,1122].eqiad.wmnet
----- OUTPUT for command #1: 'cat /sys/devices..._governor | uniq' -----
schedutil

Creating this ticket to:

  • Enable performance governor on all master-eligible hosts
  • Observe changes (if any) during the cluster restart scheduled for tomorrow (ref T324335).

Note that non-masters are also responsible for reporting their status to the cluster, so it's possible we might have to enable the performance governor on all hosts. We'll reach out to DC Ops before making a change like that, as it could impact power draw.

Event Timeline

Here's a breakdown of master-eligibles by row. These changes will likely raise power usage in the DC, so it's important we keep DC Ops in the loop. The checked hosts already have the performance governor enabled.


EQIAD:

Row A:

  • cirrussearch1068.eqiad.wmnet
  • cirrussearch1073.eqiad.wmnet

Row B:

  • cirrussearch1074.eqiad.wmnet
  • cirrussearch1075.eqiad.wmnet
  • cirrussearch1076.eqiad.wmnet

Row C:

  • cirrussearch1081.eqiad.wmnet
  • cirrussearch1118.eqiad.wmnet
  • cirrussearch1119.eqiad.wmnet

Row D:

  • cirrussearch1122.eqiad.wmnet

Row E:

  • cirrussearch1093.eqiad.wmnet
  • cirrussearch1094.eqiad.wmnet
  • cirrussearch1095.eqiad.wmnet

Row F:

  • cirrussearch1098.eqiad.wmnet
  • cirrussearch1100.eqiad.wmnet
  • cirrussearch1102.eqiad.wmnet

CODFW:

Row A:

  • cirrussearch2061.codfw.wmnet
  • cirrussearch2073.codfw.wmnet
  • cirrussearch2076.codfw.wmnet
  • cirrussearch2111.codfw.wmnet

Row B:

  • cirrussearch2080.codfw.wmnet
  • cirrussearch2092.codfw.wmnet
  • cirrussearch2093.codfw.wmnet

Row C:

  • cirrussearch2081.codfw.wmnet
  • cirrussearch2083.codfw.wmnet
  • cirrussearch2100.codfw.wmnet

Row D:

  • cirrussearch2084.codfw.wmnet
  • cirrussearch2086.codfw.wmnet
  • cirrussearch2106.codfw.wmnet
  • cirrussearch2108.codfw.wmnet
  • cirrussearch2115.codfw.wmnet

Change #1321102 had a related patch set uploaded (by Bking; author: Bking):

[operations/puppet@production] cirrussearch: Enable performance governor on all master-eligible hosts

https://gerrit.wikimedia.org/r/1321102

Change #1321102 merged by Bking:

[operations/puppet@production] cirrussearch: Enable performance governor on master-eligible hosts

https://gerrit.wikimedia.org/r/1321102

Per the above change, I enabled the performance governor for all EQIAD master-eligible hosts. Unfortunately, this did not prevent the cluster quorum issues, so I'm going to go ahead and revert the change before I forget about it.

The above change has been reverted. Since it doesn't seem to be the root cause (likely resolution mentioned here ), I'm going to close out this ticket.