| Message ID | 20201126005603.1293012-1-shakeelb@google.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | memcg: add pagetable comsumption to memory.stat | expand |
On Wed, Nov 25, 2020 at 04:56:01PM -0800, Shakeel Butt wrote: > Many workloads consumes significant amount of memory in pagetables. This > patch series exposes the pagetable comsumption for each memory cgroup. Hi Shakeel! The code looks good to me. However I'm not sure I understand what's the use case for the new statistics? Can you, please, elaborate a bit more here? From a very first glance, the size of pagetables should be _roughly_ equal to the size_of(pte)/PAGE_SIZE*(size of a cgroup) and should not exceed 1% of the cgroup size. So for all but very large cgroups the value will be in the noise of per-cpu counters. Perhaps I'm missing some important cases? Thanks! > > Shakeel Butt (2): > mm: move lruvec stats update functions to vmstat.h > mm: memcontrol: account pagetables per node > > Documentation/admin-guide/cgroup-v2.rst | 3 + > arch/nds32/mm/mm-nds32.c | 6 +- > drivers/base/node.c | 2 +- > fs/proc/meminfo.c | 2 +- > include/linux/memcontrol.h | 112 ------------------------ > include/linux/mm.h | 11 ++- > include/linux/mmzone.h | 2 +- > include/linux/vmstat.h | 104 ++++++++++++++++++++++ > mm/memcontrol.c | 19 ++++ > mm/page_alloc.c | 6 +- > 10 files changed, 142 insertions(+), 125 deletions(-) > > -- > 2.29.2.454.gaff20da3a2-goog >
On Mon, Nov 30, 2020 at 12:34 PM Roman Gushchin <guro@fb.com> wrote: > > On Wed, Nov 25, 2020 at 04:56:01PM -0800, Shakeel Butt wrote: > > Many workloads consumes significant amount of memory in pagetables. This > > patch series exposes the pagetable comsumption for each memory cgroup. > > Hi Shakeel! > > The code looks good to me. However I'm not sure I understand what's the > use case for the new statistics? Can you, please, elaborate a bit more here? > > From a very first glance, the size of pagetables should be _roughly_ equal > to the size_of(pte)/PAGE_SIZE*(size of a cgroup) and should not exceed 1% > of the cgroup size. So for all but very large cgroups the value will be > in the noise of per-cpu counters. Perhaps I'm missing some important cases? > I think this is in general a good metric to have but one specific use-case we have is the user space network driver which mmaps the memory of the applications for zero copy data transfers. This driver can consume a large amount of memory in page tables. So, this metric becomes really useful here. Shakeel
On Mon, Nov 30, 2020 at 01:01:18PM -0800, Shakeel Butt wrote: > On Mon, Nov 30, 2020 at 12:34 PM Roman Gushchin <guro@fb.com> wrote: > > > > On Wed, Nov 25, 2020 at 04:56:01PM -0800, Shakeel Butt wrote: > > > Many workloads consumes significant amount of memory in pagetables. This > > > patch series exposes the pagetable comsumption for each memory cgroup. > > > > Hi Shakeel! > > > > The code looks good to me. However I'm not sure I understand what's the > > use case for the new statistics? Can you, please, elaborate a bit more here? > > > > From a very first glance, the size of pagetables should be _roughly_ equal > > to the size_of(pte)/PAGE_SIZE*(size of a cgroup) and should not exceed 1% > > of the cgroup size. So for all but very large cgroups the value will be > > in the noise of per-cpu counters. Perhaps I'm missing some important cases? > > > > I think this is in general a good metric to have but one specific > use-case we have is the user space network driver which mmaps the > memory of the applications for zero copy data transfers. This driver > can consume a large amount of memory in page tables. So, this metric > becomes really useful here. Got it, thank you for the explanation! Would you mind to add this text to the cover letter as an example? Acked-by: Roman Gushchin <guro@fb.com> for the series. Thanks!
On Mon, Nov 30, 2020 at 1:10 PM Roman Gushchin <guro@fb.com> wrote: > > On Mon, Nov 30, 2020 at 01:01:18PM -0800, Shakeel Butt wrote: > > On Mon, Nov 30, 2020 at 12:34 PM Roman Gushchin <guro@fb.com> wrote: > > > > > > On Wed, Nov 25, 2020 at 04:56:01PM -0800, Shakeel Butt wrote: > > > > Many workloads consumes significant amount of memory in pagetables. This > > > > patch series exposes the pagetable comsumption for each memory cgroup. > > > > > > Hi Shakeel! > > > > > > The code looks good to me. However I'm not sure I understand what's the > > > use case for the new statistics? Can you, please, elaborate a bit more here? > > > > > > From a very first glance, the size of pagetables should be _roughly_ equal > > > to the size_of(pte)/PAGE_SIZE*(size of a cgroup) and should not exceed 1% > > > of the cgroup size. So for all but very large cgroups the value will be > > > in the noise of per-cpu counters. Perhaps I'm missing some important cases? > > > > > > > I think this is in general a good metric to have but one specific > > use-case we have is the user space network driver which mmaps the > > memory of the applications for zero copy data transfers. This driver > > can consume a large amount of memory in page tables. So, this metric > > becomes really useful here. > > Got it, thank you for the explanation! > Would you mind to add this text to the cover letter as an example? Sure. > > Acked-by: Roman Gushchin <guro@fb.com> > for the series. Thanks for the review.