[ECS] [request]: Support For Restricting Containers To A Single NUMA ...

Trang chủ » C5 9xl » [ECS] [request]: Support For Restricting Containers To A Single NUMA ...

Có thể bạn quan tâm

Skip to content Dismiss alert {{ message }} / containers-roadmap Public
  • Notifications You must be signed in to change notification settings
  • Fork 330
  • Star 5.3k
  • Code
  • Issues 1.5k
  • Pull requests
  • Actions
  • Projects 1
  • Security

    Uh oh!

    There was an error while loading. Please reload this page.

  • Insights
Additional navigation options [ECS] [request]: Support for restricting containers to a single NUMA node #1093New issueNew issueOpenOpen[ECS] [request]: Support for restricting containers to a single NUMA node#1093LabelsProposedCommunity submitted issue@tomelliff

Description

@tomellifftomelliffopened on Sep 30, 2020

Community Note

  • Please vote on this issue by adding a 👍 reaction to the original issue to help the community and maintainers prioritize this request
  • Please do not leave "+1" or "me too" comments, they generate extra noise for issue followers and do not help prioritize the request
  • If you are interested in working on this issue or have submitted a pull request, please leave a comment

Tell us about your request Docker allows us to specify cpuset-cpus to pin to specific CPUs or more widely to use cpuset-mems to pin to specific NUMA nodes and prevent the container trying to use memory across another socket which can cause massive performance hits. I'd like this to be exposable by the ECS API but also to have a higher level requirement where we can just state the number of NUMA nodes it can use (which we would like to stick as 1). This higher level feature would then mean it's much easier to just throw any CPU heavy workload at a dual socket (or more) instance and not have to worry about NUMA issues.

Which service(s) is this request for? ECS

Tell us about the problem you're trying to solve. What are you trying to do, and why is it hard? We have some CPU heavy tasks that require a lot of memory access. When we run them on a c5.18xl or c5.24xl instance then they are much slower than when we run on c5.9xl or c5.12xl instances due to NUMA issues where computation is limited by memory access times when going across cores.

Are you currently working around this issue? We're using single socket nodes only. We'd like to be able to be flexible and use autoscaling groups that support mixed instance types including the dual socket nodes.

👍React with 👍4

Metadata

Metadata

Assignees

No one assigned

Labels

ProposedCommunity submitted issue

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions

You can’t perform that action at this time.

Từ khóa » C5 9xl