On Tue 23-08-22 20:46:43, Liu Shixin wrote:
> On 2022/8/23 15:50, Michal Hocko wrote:
> > On Mon 22-08-22 14:12:07, Andrew Morton wrote:
> >> On Mon, 22 Aug 2022 11:33:54 +0800 Liu Shixin <[email protected]> wrote:
> >>
> >>> The page on pcplist could be used, but not counted into memory free or
> >>> avaliable, and pcp_free is only showed by show_mem() for now. Since commit
> >>> d8a759b57035 ("mm, page_alloc: double zone's batchsize"), there is a
> >>> significant decrease in the display of free memory, with a large number
> >>> of cpus and zones, the number of pages in the percpu list can be very
> >>> large, so it is better to let user to know the pcp count.
> >>>
> >>> On a machine with 3 zones and 72 CPUs. Before commit d8a759b57035, the
> >>> maximum amount of pages in the pcp lists was theoretically 162MB(3*72*768KB).
> >>> After the patch, the lists can hold 324MB. It has been observed to be 114MB
> >>> in the idle state after system startup in practice(increased 80 MB).
> >>>
> >> Seems reasonable.
> > I have asked in the previous incarnation of the patch but haven't really
> > received any answer[1]. Is this a _real_ problem? The absolute amount of
> > memory could be perceived as a lot but is this really noticeable wrt
> > overall memory on those systems?
>
> This may not obvious when the memory is sufficient. However, as products monitor the
> memory to plan it. The change has caused warning.
Is it possible that the said monitor is over sensitive and looking at
wrong numbers? Overall free memory doesn't really tell much TBH.
MemAvailable is a very rough estimation as well.
In reality what really matters much more is whether the memory is
readily available when it is required and none of MemFree/MemAvailable
gives you that information in general case.
> We have also considered using /proc/zoneinfo to calculate the total
> number of pcplists. However, we think it is more appropriate to add
> the total number of pcplists to free and available pages. After all,
> this part is also free pages.
Those free pages are not generally available as exaplained. They are
available to a specific CPU, drained under memory pressure and other
events but still there is no guarantee a specific process can harvest
that memory because the pcp caches are replenished all the time.
So in a sense it is a semi-hidden memory.
That being said, I am still not convinced this is actually going to help
all that much. You will see a slightly different numbers which do not
tell much one way or another and if the sole reason for tweaking these
numbers is that some monitor is complaining because X became X-epsilon
then this sounds like a weak justification to me. That epsilon happens
all the time because there are quite some hidden caches that are
released under memory pressure. I am not sure it is maintainable to
consider each one of them and pretend that MemFree/MemAvailable is
somehow precise. It has never been and likely never will be.
--
Michal Hocko
SUSE Labs
On Tue, Aug 23, 2022 at 03:37:52PM +0200, Michal Hocko wrote:
> On Tue 23-08-22 20:46:43, Liu Shixin wrote:
> > On 2022/8/23 15:50, Michal Hocko wrote:
> > > On Mon 22-08-22 14:12:07, Andrew Morton wrote:
> > >> On Mon, 22 Aug 2022 11:33:54 +0800 Liu Shixin <[email protected]> wrote:
> > >>
> > >>> The page on pcplist could be used, but not counted into memory free or
> > >>> avaliable, and pcp_free is only showed by show_mem() for now. Since commit
> > >>> d8a759b57035 ("mm, page_alloc: double zone's batchsize"), there is a
> > >>> significant decrease in the display of free memory, with a large number
> > >>> of cpus and zones, the number of pages in the percpu list can be very
> > >>> large, so it is better to let user to know the pcp count.
> > >>>
> > >>> On a machine with 3 zones and 72 CPUs. Before commit d8a759b57035, the
> > >>> maximum amount of pages in the pcp lists was theoretically 162MB(3*72*768KB).
> > >>> After the patch, the lists can hold 324MB. It has been observed to be 114MB
> > >>> in the idle state after system startup in practice(increased 80 MB).
> > >>>
> > >> Seems reasonable.
> > > I have asked in the previous incarnation of the patch but haven't really
> > > received any answer[1]. Is this a _real_ problem? The absolute amount of
> > > memory could be perceived as a lot but is this really noticeable wrt
> > > overall memory on those systems?
Let me provide some other numbers, from the desktop side. On a low-end
chromebook with 4GB RAM and a dual-core CPU, after commit b92ca18e8ca5
(mm/page_alloc: disassociate the pcp->high from pcp->batch) the max
amount of PCP pages increased 56x times: from 2.9MB (1.45 per CPU) to
165MB (82.5MB per CPU).
On such a system, memory pressure conditions are not a rare occurrence,
so several dozen MB make a lot of difference.
(The reason it increased so much is because it now corresponds to the
low watermark, which is 165MB. And the low watermark, in turn, is so
high because of khugepaged, which bumps up min_free_kbytes to 132MB
regardless of the total amount of memory.)
> > This may not obvious when the memory is sufficient. However, as products monitor the
> > memory to plan it. The change has caused warning.
>
> Is it possible that the said monitor is over sensitive and looking at
> wrong numbers? Overall free memory doesn't really tell much TBH.
> MemAvailable is a very rough estimation as well.
>
> In reality what really matters much more is whether the memory is
> readily available when it is required and none of MemFree/MemAvailable
> gives you that information in general case.
>
> > We have also considered using /proc/zoneinfo to calculate the total
> > number of pcplists. However, we think it is more appropriate to add
> > the total number of pcplists to free and available pages. After all,
> > this part is also free pages.
>
> Those free pages are not generally available as exaplained. They are
> available to a specific CPU, drained under memory pressure and other
> events but still there is no guarantee a specific process can harvest
> that memory because the pcp caches are replenished all the time.
> So in a sense it is a semi-hidden memory.
I was intuitively assuming that per-CPU pages should be always available
for allocation without resorting to paging out allocated pages (and thus
it should be non-controversially a good idea to include per-CPU pages in
MemFree, to make it more accurate).
But looking at the code in __alloc_pages() and around, I see you are
right: we don't try draining other CPUs' PCP lists *before* resorting to
direct reclaim, compaction etc.
BTW, why not? Shouldn't draining PCP lists be cheaper than pageout() in
any case?
> That being said, I am still not convinced this is actually going to help
> all that much. You will see a slightly different numbers which do not
> tell much one way or another and if the sole reason for tweaking these
> numbers is that some monitor is complaining because X became X-epsilon
> then this sounds like a weak justification to me. That epsilon happens
> all the time because there are quite some hidden caches that are
> released under memory pressure. I am not sure it is maintainable to
> consider each one of them and pretend that MemFree/MemAvailable is
> somehow precise. It has never been and likely never will be.
> --
> Michal Hocko
> SUSE Labs
On Fri 24-11-23 18:54:54, Dmytro Maluka wrote:
[...]
> But looking at the code in __alloc_pages() and around, I see you are
> right: we don't try draining other CPUs' PCP lists *before* resorting to
> direct reclaim, compaction etc.
>
> BTW, why not? Shouldn't draining PCP lists be cheaper than pageout() in
> any case?
My guess would be that draining remote pcp caches is quite expensive on
its own. This requires IPIs, preempting whatever is running there and
wait for the all the cpus with pcp caches to be done. On the other hand
reclaiming a mostly clean page cache could be much less expensive.
Also consider that refilling those pcp caches is not free either (you
might hit zone lock contetion and who knows what else).
Last but not least also consider that many systems could be just on the
edge of low/min watermark with a lot of cached data. If we drained all
pcp caches whenever we reclaim this could just make the cache pointless.
All that being said, I do not remember any actual numbers or research
about this.
--
Michal Hocko
SUSE Labs