On Thu 22-09-22 12:18:04, David Rientjes wrote: > On Wed, 21 Sep 2022, Gang Li wrote: > > > cpuset confine processes to processor and memory node subsets. > > When a process in cpuset triggers oom, it may kill a completely > > irrelevant process on another numa node, which will not release any > > memory for this cpuset. > > > > It seems that `CONSTRAINT_CPUSET` is not really doing much these > > days. Using CONSTRAINT_CPUSET, we can easily achieve node aware oom > > killing by selecting victim from the cpuset which triggers oom. > > > > Suggested-by: Michal Hocko <mhocko@xxxxxxxx> > > Signed-off-by: Gang Li <ligang.bdlg@xxxxxxxxxxxxx> > > Hmm, is this the right approach? > > If a cpuset results in a oom condition, is there a reason why we'd need to > find a process from within that cpuset to kill? I think the idea is to > free memory on the oom set of nodes (cpuset.mems) and that can happen by > killing a process that is not a member of this cpuset. I would argue that the current cpuset should be considered first because chances are that it will already have the biggest memory consumption from the constrained NUMA nodes. At least that would be the case when cpusets are used to partition the system into exclusive NUMA domains. Situation gets more complex with overlapping nodemasks in different cpusets but I believe our existing semantic sucks already for those usecases already because we just shoot a random process with an unknown amount of memory allocated from the constrained nodemask. This new semantic is not much worse. We could find a real oom victim under a different cpuset but the current semantic could as well kill a large memory consumer with a tiny footprint on the target node. With the cpuset view the potential damage is more targeted in many cases. > I understand the challenges of creating a NUMA aware oom killer to target > memory that is actually resident on an oom node, but this approach doesn't > seem right and could actually lead to pathological cases where a small > process trying to fork in an otherwise empty cpuset is repeatedly oom > killing when we'd actually prefer to kill a single large process. Yeah, that is possible and something to consider. One way to go about that is to make the selection from all cpusets with an overlap with the requested nodemask (probably with a preference to more constrained ones). In any case let's keep in mind that this is a mere heuristic. We just need to kill some process, it is not really feasible to aim for the best selection. We should just try to reduce the harm. Our exisiting cpuset based OOM is effectivelly random without any clear relation to cpusets so I would be open to experimenting in this area. -- Michal Hocko SUSE Labs
