On Wed 21-04-21 06:57:43, Shakeel Butt wrote: > On Wed, Apr 21, 2021 at 12:16 AM Michal Hocko <mhocko@xxxxxxxx> wrote: > > > [...] > > > To decide when to kill, the oom-killer has to read a lot of metrics. > > > It has to open a lot of files to read them and there will definitely > > > be new allocations involved in those operations. For example reading > > > memory.stat does a page size allocation. Similarly, to perform action > > > the oom-killer may have to read cgroup.procs file which again has > > > allocation inside it. > > > > True but many of those can be avoided by opening the file early. At > > least seq_file based ones will not allocate later if the output size > > doesn't increase. Which should be the case for many. I think it is a > > general improvement to push those who allocate during read to an open > > time allocation. > > > > I agree that this would be a general improvement but it is not always > possible (see below). It would be still great to invest into those improvements. And I would be really grateful to learn about bottlenecks from the existing kernel interfaces you have found on the way. > > > Regarding sophisticated oom policy, I can give one example of our > > > cluster level policy. For robustness, many user facing jobs run a lot > > > of instances in a cluster to handle failures. Such jobs are tolerant > > > to some amount of failures but they still have requirements to not let > > > the number of running instances below some threshold. Normally killing > > > such jobs is fine but we do want to make sure that we do not violate > > > their cluster level agreement. So, the userspace oom-killer may > > > dynamically need to confirm if such a job can be killed. > > > > What kind of data do you need to examine to make those decisions? > > > > Most of the time the cluster level scheduler pushes the information to > the node controller which transfers that information to the > oom-killer. However based on the freshness of the information the > oom-killer might request to pull the latest information (IPC and RPC). I cannot imagine any OOM handler to be reliable if it has to depend on other userspace component with a lower resource priority. OOM handlers are fundamentally complex components which has to reduce their dependencies to the bare minimum. > [...] > > > > > > I was thinking of simply prctl(SET_MEMPOOL, bytes) to assign mempool > > > to a thread (not shared between threads) and prctl(RESET_MEMPOOL) to > > > free the mempool. > > > > I am not a great fan of prctl. It has become a dumping ground for all > > mix of unrelated functionality. But let's say this is a minor detail at > > this stage. > > I agree this does not have to be prctl(). > > > So you are proposing to have a per mm mem pool that would be > > I was thinking of per-task_struct instead of per-mm_struct just for simplicity. > > > used as a fallback for an allocation which cannot make a forward > > progress, right? > > Correct > > > Would that pool be preallocated and sitting idle? > > Correct > > > What kind of allocations would be allowed to use the pool? > > I was thinking of any type of allocation from the oom-killer (or > specific threads). Please note that the mempool is the backup and only > used in the slowpath. > > > What if the pool is depleted? > > This would mean that either the estimate of mempool size is bad or > oom-killer is buggy and leaking memory. > > I am open to any design directions for mempool or some other way where > we can provide a notion of memory guarantee to oom-killer. OK, thanks for clarification. There will certainly be hard problems to sort out[1] but the overall idea makes sense to me and it sounds like a much better approach than a OOM specific solution. [1] - how the pool is going to be replenished without hitting all potential reclaim problems (thus dependencies on other all tasks directly/indirectly) yet to not rely on any background workers to do that on the task behalf without a proper accounting etc... -- Michal Hocko SUSE Labs
