Re: [PATCH 0/6 v2] Calculate pcp->high based on zone sizes and active CPUs

On 5/28/21 10:55 AM, Mel Gorman wrote:

On Thu, May 27, 2021 at 12:36:21PM -0700, Dave Hansen wrote:

quoted

Hi Mel,

Feng Tang tossed these on a "Cascade Lake" system with 96 threads and
~512G of persistent memory and 128G of DRAM.  The PMEM is in "volatile
use" mode and being managed via the buddy just like the normal RAM.

The PMEM zones are big ones:

        present  65011712 = 248 G
        high       134595 = 525 M

The PMEM nodes, of course, don't have any CPUs in them.

With your series, the pcp->high value per-cpu is 69584 pages or about
270MB per CPU.  Scaled up by the 96 CPU threads, that's ~26GB of
worst-case memory in the pcps per zone, or roughly 10% of the size of
the zone.

I did see quite a few pcp->counts above 60,000, so it's definitely
possible in practice to see the pcps filled up.  This was not observed
to cause any actual problems in practice.  But, it's still a bit worrisome.

Ok, it does have the potential to trigger early reclaim as pages are
stored on remote PCP lists. The problem would be transient because
vmstat would drain those pages over time but still, how about this patch
on top of the series?

--8<--
mm/page_alloc: Split pcp->high across all online CPUs for cpuless nodes

Dave Hansen reported the following about Feng Tang's tests on a machine
with persistent memory onlined as a DRAM-like device.

  Feng Tang tossed these on a "Cascade Lake" system with 96 threads and
  ~512G of persistent memory and 128G of DRAM.  The PMEM is in "volatile
  use" mode and being managed via the buddy just like the normal RAM.

  The PMEM zones are big ones:

        present  65011712 = 248 G
        high       134595 = 525 M

  The PMEM nodes, of course, don't have any CPUs in them.

  With your series, the pcp->high value per-cpu is 69584 pages or about
  270MB per CPU.  Scaled up by the 96 CPU threads, that's ~26GB of
  worst-case memory in the pcps per zone, or roughly 10% of the size of
  the zone.

This should not cause a problem as such although it could trigger reclaim
due to pages being stored on per-cpu lists for CPUs remote to a node. It
is not possible to treat cpuless nodes exactly the same as normal nodes
but the worst-case scenario can be mitigated by splitting pcp->high across
all online CPUs for cpuless memory nodes.

Suggested-by: Dave Hansen <redacted>
Signed-off-by: Mel Gorman <redacted>

Acked-by: Vlastimil Babka <redacted>

Maybe we should even consider distinguishing high limits for local-to-cpu zones
vs remote, for example for the local-to-cpu zones we would divide by the number
of local cpus, for remote-to-cpu zones we would divide by all cpus.

Because we can expect cpus to allocate mostly from local zones, so leaving more
pages on percpu for those zones can be beneficial.

But as the motivation here was to reduce lock contention on freeing, that's less
clear. We probably can't expect the cpu to be freeing mostly local pages (in
case of e.g. a large process exiting), because no mechanism works towards that,
or does it? In case of cpu freeing to remote zone, the lower high limit could hurt.

So that would have to be evaluated if that works in practice. Out of scope here,
just an idea to discuss.