Details
-
Bug
-
Resolution: Unresolved
-
Normal
-
None
-
4.13.z, 4.14.z, 4.15.z, 4.16.0
-
No
-
False
-
Description
Description of problem:
We deploy /etc/sysctl.d/inotify.conf which was meant to increase the number of watches to 65k over the 8k default present in RHEL8. However RHEL9 kernel moved to a model where the value scales based on the amount of memory present in the system using up to 1% of addressable memory. See https://github.com/torvalds/linux/commit/92890123749bafc317bbfacbe0a62ce08d78efb7 and https://github.com/openshift/machine-config-operator/pull/2944#issuecomment-1030236808 In RHCOS9 now it's no longer necessary and the default provides better scalability
Version-Release number of selected component (if applicable):
4.13.z, 4.14.z, 4.15.z, 4.16.0
How reproducible:
Install 4.13 or later
Steps to Reproduce:
1. sysctl fs.inotify.max_user_watches 2. 3.
Actual results:
observe 65k value independent of instance size
Expected results:
value scaled based on memory up to 1048576 on instances with 128GiB or more
Additional info:
I did a quick test on small instances down to 4GiB to ensure that we don't scale down below the 65k value in supported instances, at least in my quick test that didn't happen. We may also want to consult with the Node Tuning Operator folks as it looks like the previous attempt to remove this file may've interacted with them. Finally, I guess since we still support RHEL Package Based workers (formerly known as RHEL workers) we likely need to find some way to ensure that we still apply these values in RHEL8. I don't know how to do that via templates. Apparently other platforms like EKS have a value of 120,000.
CC jmencak