On Tue, Mar 12, 2019 at 1:05 AM Michal Hocko <mhocko@xxxxxxxxxx> wrote: > > On Mon 11-03-19 15:15:35, Suren Baghdasaryan wrote: > > On Mon, Mar 11, 2019 at 1:46 PM Sultan Alsawaf <sultan@xxxxxxxxxxxxxxx> wrote: > > > > > > On Mon, Mar 11, 2019 at 01:10:36PM -0700, Suren Baghdasaryan wrote: > > > > The idea seems interesting although I need to think about this a bit > > > > more. Killing processes based on failed page allocation might backfire > > > > during transient spikes in memory usage. > > > > > > This issue could be alleviated if tasks could be killed and have their pages > > > reaped faster. Currently, Linux takes a _very_ long time to free a task's memory > > > after an initial privileged SIGKILL is sent to a task, even with the task's > > > priority being set to the highest possible (so unwanted scheduler preemption > > > starving dying tasks of CPU time is not the issue at play here). I've > > > frequently measured the difference in time between when a SIGKILL is sent for a > > > task and when free_task() is called for that task to be hundreds of > > > milliseconds, which is incredibly long. AFAIK, this is a problem that LMKD > > > suffers from as well, and perhaps any OOM killer implementation in Linux, since > > > you cannot evaluate effect you've had on memory pressure by killing a process > > > for at least several tens of milliseconds. > > > > Yeah, killing speed is a well-known problem which we are considering > > in LMKD. For example the recent LMKD change to assign process being > > killed to a cpuset cgroup containing big cores cuts the kill time > > considerably. This is not ideal and we are thinking about better ways > > to expedite the cleanup process. > > If you design is relies on the speed of killing then it is fundamentally > flawed AFAICT. You cannot assume anything about how quickly a task dies. > It might be blocked in an uninterruptible sleep or performin an > operation which takes some time. Sure, oom_reaper might help here but > still. That's what I was considering. This is not a silver bullet but increased speed would not hurt. > The only way to control the OOM behavior pro-actively is to throttle > allocation speed. We have memcg high limit for that purpose. Along with > PSI, I can imagine a reasonably working user space early oom > notifications and reasonable acting upon that. That makes sense and we are working in this direction. > -- > Michal Hocko > SUSE Labs Thanks, Suren.
