Re: [RFC PATCH] memcg: expose children memory usage for root

["T.J. Mercier" <tjmercier@google.com>]

On Fri, Jul 26, 2024 at 9:26 AM Yosry Ahmed <yosryahmed@google.com> wrote:
>
> On Fri, Jul 26, 2024 at 8:48 AM Shakeel Butt <shakeel.butt@linux.dev> wrote:
> >
> > On Thu, Jul 25, 2024 at 04:20:45PM GMT, Yosry Ahmed wrote:
> > > On Mon, Jul 22, 2024 at 3:53 PM Shakeel Butt <shakeel.butt@linux.dev> wrote:
> > > >
> > > > Linux kernel does not expose memory.current on the root memcg and there
> > > > are applications which have to traverse all the top level memcgs to
> > > > calculate the total memory charged in the system. This is more expensive
> > > > (directory traversal and multiple open and reads) and is racy on a busy
> > > > machine. As the kernel already have the needed information i.e. root's
> > > > memory.current, why not expose that?
> > > >
> > > > However root's memory.current will have a different semantics than the
> > > > non-root's memory.current as the kernel skips the charging for root, so
> > > > maybe it is better to have a different named interface for the root.
> > > > Something like memory.children_usage only for root memcg.
> > > >
> > > > Now there is still a question that why the kernel does not expose
> > > > memory.current for the root. The historical reason was that the memcg
> > > > charging was expensice and to provide the users to bypass the memcg
> > > > charging by letting them run in the root. However do we still want to
> > > > have this exception today? What is stopping us to start charging the
> > > > root memcg as well. Of course the root will not have limits but the
> > > > allocations will go through memcg charging and then the memory.current
> > > > of root and non-root will have the same semantics.
> > > >
> > > > This is an RFC to start a discussion on memcg charging for root.
> > >
> > > I vaguely remember when running some netperf tests (tcp_rr?) in a
> > > cgroup that the performance decreases considerably with every level
> > > down the hierarchy. I am assuming that charging was a part of the
> > > reason. If that's the case, charging the root will be similar to
> > > moving all workloads one level down the hierarchy in terms of charging
> > > overhead.
> >
> > No, the workloads running in non-root memcgs will not see any
> > difference. Only the workloads running in root will see charging
> > overhead.
>
> Oh yeah we already charge the root's page counters hierarchically in
> the upstream kernel, we just do not charge them if the origin of the
> charge is the root itself.
>
> We also have workloads that iterate top-level memcgs to calculate the
> total charged memory, so memory.children_usage for the root memcg
> would help.
>
> As for memory.current, do you have any data about how much memory is
> charged to the root itself?

Yeah I wonder if we'd be able to see any significant regressions for
stuff that lives there today if we were to start charging it. I can
try running a test with Android next week. I guess try_charge() is the
main thing that would need to change to allow root charges?

> We think of the memory charged to the root
> as system overhead, while the memory charged to top-level memcgs
> isn't.
>
> So basically total_memory - root::memory.children_usage would be a
> fast way to get a rough estimation of system overhead. The same would
> not apply for total_memory - root::memory.current if I understand
> correctly.
>

On Fri, Jul 26, 2024 at 9:26 AM Yosry Ahmed <yosryahmed@google.com> wrote:
>
> On Fri, Jul 26, 2024 at 8:48 AM Shakeel Butt <shakeel.butt@linux.dev> wrote:
> >
> > On Thu, Jul 25, 2024 at 04:20:45PM GMT, Yosry Ahmed wrote:
> > > On Mon, Jul 22, 2024 at 3:53 PM Shakeel Butt <shakeel.butt@linux.dev> wrote:
> > > >
> > > > Linux kernel does not expose memory.current on the root memcg and there
> > > > are applications which have to traverse all the top level memcgs to
> > > > calculate the total memory charged in the system. This is more expensive
> > > > (directory traversal and multiple open and reads) and is racy on a busy
> > > > machine. As the kernel already have the needed information i.e. root's
> > > > memory.current, why not expose that?
> > > >
> > > > However root's memory.current will have a different semantics than the
> > > > non-root's memory.current as the kernel skips the charging for root, so
> > > > maybe it is better to have a different named interface for the root.
> > > > Something like memory.children_usage only for root memcg.
> > > >
> > > > Now there is still a question that why the kernel does not expose
> > > > memory.current for the root. The historical reason was that the memcg
> > > > charging was expensice and to provide the users to bypass the memcg
> > > > charging by letting them run in the root. However do we still want to
> > > > have this exception today? What is stopping us to start charging the
> > > > root memcg as well. Of course the root will not have limits but the
> > > > allocations will go through memcg charging and then the memory.current
> > > > of root and non-root will have the same semantics.
> > > >
> > > > This is an RFC to start a discussion on memcg charging for root.
> > >
> > > I vaguely remember when running some netperf tests (tcp_rr?) in a
> > > cgroup that the performance decreases considerably with every level
> > > down the hierarchy. I am assuming that charging was a part of the
> > > reason. If that's the case, charging the root will be similar to
> > > moving all workloads one level down the hierarchy in terms of charging
> > > overhead.
> >
> > No, the workloads running in non-root memcgs will not see any
> > difference. Only the workloads running in root will see charging
> > overhead.
>
> Oh yeah we already charge the root's page counters hierarchically in
> the upstream kernel, we just do not charge them if the origin of the
> charge is the root itself.
>
> We also have workloads that iterate top-level memcgs to calculate the
> total charged memory, so memory.children_usage for the root memcg
> would help.
>
> As for memory.current, do you have any data about how much memory is
> charged to the root itself? We think of the memory charged to the root
> as system overhead, while the memory charged to top-level memcgs
> isn't.
>
> So basically total_memory - root::memory.children_usage would be a
> fast way to get a rough estimation of system overhead. The same would
> not apply for total_memory - root::memory.current if I understand
> correctly.
>