Red Hat Bugzilla – Bug 1463489
[RFE] make it possible to install Satellite's httpd with mpm_event
Last modified: 2018-05-18 01:48:56 EDT
Description of problem:
We did some measurement for number of concurrent registrations with Satellite on default mpm_prefork vs. mpm_event and with mpm_event, we saw significant saving in RSS memory consumed by httpd processes.
Version-Release number of selected component (if applicable):
Steps to Reproduce:
1. measure RSS usage of all httpd processes together with default (currently
mpm_prefork) setting during concurrent registration of 140 hosts
2. measure RSS usage of all httpd processes together with mpm_event setting
configured during concurrent registration of 140 hosts
RSS memory usage in first case was about 2 GB, in second case 760 MB.
As discussed, it is important to carefully sanity-test Satellite in regards of functionality and performance.
Created attachment 1289934 [details]
httpd RSS memory usage comparision for different MPM setting and different number of concurrent registrations
Graph shows memory consumption on Satellite in a time
period of all the clients being processed (and returned 0 - except for a
last row where some of them failed because of timeout).
So if registration command on clients was running:
client1: 12:00:00 - 12:01:00
client2: 12:00:00 - 12:01:10
client3: 12:00:20 - 12:01:25
graph shows time range of 12:00:00 - 12:01:25 (well, actually it adds 10
second boundary, so it shows 11:59:50 - 12:01:35).
Created attachment 1289935 [details]
same comparison of graph but for system load
can we clone this bz to upstream foreman issue
The easiest way is adding a line to custom-hiera.yaml:
Then rerunning the installer should be sufficient to switch the MPM worker and should work for any installer version that supports custom-hiera.yaml.
Created redmine issue http://projects.theforeman.org/issues/20889 from this bug
Moving this bug to POST for triage into Satellite 6 since the upstream issue http://projects.theforeman.org/issues/20889 has been resolved.
It turns out there's a bug in Apache with the MPM event worker that's fixed in 2.4.25: AH00485: scoreboard is full, not at MaxRequestWorkers. Due to this we might roll back the fix in upstream until the fix is backported. Using the httpd24 SCL (which contains the fix) is not an option because its too invasive in the system.
Setting back to NEW since it was reverted in upstream. We either need a fix in EL7's httpd (which was closed as WONTFIX) or move to the httpd SCL.
Upstream bug assigned to email@example.com