On Thu, Mar 7, 2019 at 3:08 PM Patrick Donnelly <pdonnell@xxxxxxxxxx> wrote: > > On Thu, Mar 7, 2019 at 3:02 PM Sage Weil <sweil@xxxxxxxxxx> wrote: > > > > On Thu, 7 Mar 2019, Gregory Farnum wrote: > > > > With that caveat, it seems like we should *also* look at the cgroup limit > > > > * some factor (e.g., .8) as use it as a ceiling for *_memory_target... > > > > > > Even that probably has too many assumptions though. If we're in a > > > cgroup in a non-Kubernetes/plain container context, there's every > > > possibility that we aren't the only consumer of that cgroup's resource > > > limits (for instance, all of the Ceph processes on a machine in one > > > cgroup, or vhost-style process isolation with all of a tenant in one > > > cgroup), so there would need to be another input telling us how much > > > memory to target for any individual process. :/ > > > > That's good point. We could use it as a *ceiling*, though... > > In practice, I don't think anyone places more than one process in a > cgroup like this. Who does this vhost-style isolation you're referring > to? I believe it's the original use case for cgroups — shared hosting providers or anybody who was doing VPS or "light virtualization" or whatever was giving each user a separate cgroup, but all their processes ran within just the one.
