[release/2.1] Determine memory load based on cgroup usage. #19650
Conversation
cgroup usage is used to trigger oom kills. It includes rss and file cache of the cgroup. The implementation was only using the process rss to determine memory load. This is less than the cgroup usage and leads to oom kills due to GC not being triggered soon enough.
Description.NET core applications are killed by OpenShift because they exceed their assigned memory. OpenShift/Kubernetes informs the app via the sysfs limit_in_bytes. Then memory is monitored by the oom killer based on sysfs usage_in_bytes. .NET Core is using /proc/self/statm instead, which includes only RSS and not memory in cache as the usage_in_bytes. Customer ImpactCustomer’s applications are killed by OpenShift because they exceed their assigned memory. Regression?No RiskNo Original issue: #19060 |
|
Thanks for fixing this! We think we've been hitting this issue in ECS, so I'll give a +1 for having it fixed in a 2.1 patch. |
jamshedd
added
Servicing-approved
|
Approved for 2.1.5 |
|
This is expected to go out in October patch. |
|
will this solve https://github.com/dotnet/coreclr/issues/18044 as well? |
|
@devKlausS it should, .NET Core now uses the same limit as the Linux OOM killer. |
|
@tmds thanks for the quick reply! I did a retest with 2.1.5 and I am afraid the issued is not solved yet. See my update https://github.com/dotnet/coreclr/issues/18044#issuecomment-426688885 If you need more details please let me know. |
Port #19518 to 2.1
CC @janvorli