On Mon, 12 Aug 2024, Kaiyang Zhao wrote: > On Sun, Aug 11, 2024 at 01:16:53PM -0700, David Rientjes wrote: > > Hi Kaiyang, have you considered per-memcg control over NUMA balancing > > operations as well? > > > > Wondering if that's the direction that you're heading in, because it would > > be very useful to be able to control NUMA balancing at memcg granularity > > on multi-tenant systems. > > > > I mentioned this at LSF/MM/BPF this year. If people believe this is out > > of scope for memcg, that would be good feedback as well. > > Yes that's exactly where we are heading -- per-cgroup control of NUMA > balancing operations in the context of memory tiering with CXL memory, > by extending the concept of memory.low and memory.high. The use case is > enabling a fair share of top tier memory across containers. > Thanks Kaiyang, that will be very useful to test out, looking forward to seeing the patches! Does this include top-tier specific memory limits as well? And is your primary motivation the promotion path through NUMA Balancing or are you also looking at demotion to develop a comprehensive policy for memory placement using these limits? > I'm collaborating with Meta on this, and we already have an implementation > and some experiments done. The patches will go out soon. If others have > thoughts on this, please chime in. > I have lots of thoughts, but not sure if the primary motivation is around promotion only here :)
