On Thu, Dec 24, 2020 at 6:08 AM Shakeel Butt [off-list ref] wrote:

On Wed, Dec 16, 2020 at 7:45 PM Muchun Song [off-list ref] wrote:

quoted

Currently we use struct per_cpu_nodestat to cache the vmstat
counters, which leads to inaccurate statistics expecially THP

*especially

Thanks.

quoted

vmstat counters. In the systems with hundreads of processors

*hundreds

Thanks.

quoted

it can be GBs of memory. For example, for a 96 CPUs system,
the threshold is the maximum number of 125. And the per cpu
counters can cache 23.4375 GB in total.

The THP page is already a form of batched addition (it will
add 512 worth of memory in one go) so skipping the batching
seems like sensible. Although every THP stats update overflows
the per-cpu counter, resorting to atomic global updates. But
it can make the statistics more accuracy for the THP vmstat
counters.

So we convert the NR_ANON_THPS account to pages. This patch
is consistent with 8f182270dfec ("mm/swap.c: flush lru pvecs
on compound page arrival"). Doing this also can make the unit
of vmstat counters more unified. Finally, the unit of the vmstat
counters are pages, kB and bytes. The B/KB suffix can tell us
that the unit is bytes or kB. The rest which is without suffix
are pages.

Signed-off-by: Muchun Song <redacted>

I agree with the motivation behind this patch but I would like to see
some performance numbers in the commit message. We might agree to pay
the price but at least we will know what exactly that cost is.

Do you have any recommendations about benchmarks?
I can do a test. Thanks very much.

-- 
Yours,
Muchun