Re: [PATCH 0/2] mm: swap: mTHP swap allocator base on swap cluster order

[Huang, Ying]

Kairui Song <ryncsn@xxxxxxxxx> writes:

> On Fri, May 31, 2024 at 10:37 AM Huang, Ying <ying.huang@xxxxxxxxx> wrote:
>>
>> Chris Li <chrisl@xxxxxxxxxx> writes:
>>
>> > On Wed, May 29, 2024 at 7:54 PM Huang, Ying <ying.huang@xxxxxxxxx> wrote:
>> >>
>> >> Chris Li <chrisl@xxxxxxxxxx> writes:
>> >>
>> >> > Hi Ying,
>> >> >
>> >> > On Wed, May 29, 2024 at 1:57 AM Huang, Ying <ying.huang@xxxxxxxxx> wrote:
>> >> >>
>> >> >> Chris Li <chrisl@xxxxxxxxxx> writes:
>> >> >>
>> >> >> > I am spinning a new version for this series to address two issues
>> >> >> > found in this series:
>> >> >> >
>> >> >> > 1) Oppo discovered a bug in the following line:
>> >> >> > +               ci = si->cluster_info + tmp;
>> >> >> > Should be "tmp / SWAPFILE_CLUSTER" instead of "tmp".
>> >> >> > That is a serious bug but trivial to fix.
>> >> >> >
>> >> >> > 2) order 0 allocation currently blindly scans swap_map disregarding
>> >> >> > the cluster->order.
>> >> >>
>> >> >> IIUC, now, we only scan swap_map[] only if
>> >> >> !list_empty(&si->free_clusters) && !list_empty(&si->nonfull_clusters[order]).
>> >> >> That is, if you doesn't run low swap free space, you will not do that.
>> >> >
>> >> > You can still swap space in order 0 clusters while order 4 runs out of
>> >> > free_cluster
>> >> > or nonfull_clusters[order]. For Android that is a common case.
>> >>
>> >> When we fail to allocate order 4, we will fallback to order 0.  Still
>> >> don't need to scan swap_map[].  But after looking at your below reply, I
>> >> realized that the swap space is almost full at most times in your cases.
>> >> Then, it's possible that we run into scanning swap_map[].
>> >> list_empty(&si->free_clusters) &&
>> >> list_empty(&si->nonfull_clusters[order]) will become true, if we put too
>> >> many clusters in si->percpu_cluster.  So, if we want to avoid to scan
>> >> swap_map[], we can stop add clusters in si->percpu_cluster when swap
>> >> space runs low.  And maybe take clusters out of si->percpu_cluster
>> >> sometimes.
>> >
>> > One idea after reading your reply. If we run out of the
>> > nonfull_cluster[order], we should be able to use other cpu's
>> > si->percpu_cluster[] as well. That is a very small win for Android,
>>
>> This will be useful in general.  The number CPU may be large, and
>> multiple orders may be used.
>>
>> > because android does not have too many cpu. We are talking about a
>> > handful of clusters, which might not justify the code complexity. It
>> > does not change the behavior that order 0 can pollut higher order.
>>
>> I have a feeling that you don't really know why swap_map[] is scanned.
>> I suggest you to do more test and tracing to find out the reason.  I
>> suspect that there are some non-full cluster collection issues.
>>
>> >> Another issue is nonfull_cluster[order1] cannot be used for
>> >> nonfull_cluster[order2].  In definition, we should not fail order 0
>> >> allocation, we need to steal nonfull_cluster[order>0] for order 0
>> >> allocation.  This can avoid to scan swap_map[] too.  This may be not
>> >> perfect, but it is the simplest first step implementation.  You can
>> >> optimize based on it further.
>> >
>> > Yes, that is listed as the limitation of this cluster order approach.
>> > Initially we need to support one order well first. We might choose
>> > what order that is, 16K or 64K folio. 4K pages are too small, 2M pages
>> > are too big. The sweet spot might be some there in between.  If we can
>> > support one order well, we can demonstrate the value of the mTHP. We
>> > can worry about other mix orders later.
>> >
>> > Do you have any suggestions for how to prevent the order 0 polluting
>> > the higher order cluster? If we allow that to happen, then it defeats
>> > the goal of being able to allocate higher order swap entries. The
>> > tricky question is we don't know how much swap space we should reserve
>> > for each order. We can always break higher order clusters to lower
>> > order, but can't do the reserves. The current patch series lets the
>> > actual usage determine the percentage of the cluster for each order.
>> > However that seems not enough for the test case Barry has. When the
>> > app gets OOM kill that is where a large swing of order 0 swap will
>> > show up and not enough higher order usage for the brief moment. The
>> > order 0 swap entry will pollute the high order cluster. We are
>> > currently debating a "knob" to be able to reserve a certain % of swap
>> > space for a certain order. Those reservations will be guaranteed and
>> > order 0 swap entry can't pollute them even when it runs out of swap
>> > space. That can make the mTHP at least usable for the Android case.
>>
>> IMO, the bottom line is that order-0 allocation is the first class
>> citizen, we must keep it optimized.  And, OOM with free swap space isn't
>> acceptable.  Please consider the policy we used for page allocation.
>>
>> > Do you see another way to protect the high order cluster polluted by
>> > lower order one?
>>
>> If we use high-order page allocation as reference, we need something
>> like compaction to guarantee high-order allocation finally.  But we are
>> too far from that.
>>
>> For specific configuration, I believe that we can get reasonable
>> high-order swap entry allocation success rate for specific use cases.
>> For example, if we only do limited maximum number order-0 swap entries
>> allocation, can we keep high-order clusters?
>
> Isn't limiting order-0 allocation breaks the bottom line that order-0
> allocation is the first class citizen, and should not fail if there is
> space?

Sorry for confusing words.  I mean limiting maximum number order-0 swap
entries allocation in workloads, instead of limiting that in kernel.

> Just my two cents...
>
> I had a try locally based on Chris's work, allowing order 0 to use
> nonfull_clusters as Ying has suggested, and starting with low order
> and increase the order until nonfull_cluster[order] is not empty, that
> way higher order is just better protected, because unless we ran out
> of free_cluster and nonfull_cluster, direct scan won't happen.
>
> More concretely, I applied the following changes, which didn't change
> the code much:
> - In scan_swap_map_try_ssd_cluster, check nonfull_cluster first, then
> free_clusters, then discard_cluster.
> - If it's order 0, also check for (int i = 0; i < SWAP_NR_ORDERS; ++i)
> nonfull_clusters[i] cluster before scan_swap_map_try_ssd_cluster
> returns false.
>
> A quick test still using the memtier test, but decreased the swap
> device size from 10G to 8g for higher pressure.
>
> Before:
> hugepages-32kB/stats/swpout:34013
> hugepages-32kB/stats/swpout_fallback:266
> hugepages-512kB/stats/swpout:0
> hugepages-512kB/stats/swpout_fallback:77
> hugepages-2048kB/stats/swpout:0
> hugepages-2048kB/stats/swpout_fallback:1
> hugepages-1024kB/stats/swpout:0
> hugepages-1024kB/stats/swpout_fallback:0
> hugepages-64kB/stats/swpout:35088
> hugepages-64kB/stats/swpout_fallback:66
> hugepages-16kB/stats/swpout:31848
> hugepages-16kB/stats/swpout_fallback:402
> hugepages-256kB/stats/swpout:390
> hugepages-256kB/stats/swpout_fallback:7244
> hugepages-128kB/stats/swpout:28573
> hugepages-128kB/stats/swpout_fallback:474
>
> After:
> hugepages-32kB/stats/swpout:31448
> hugepages-32kB/stats/swpout_fallback:3354
> hugepages-512kB/stats/swpout:30
> hugepages-512kB/stats/swpout_fallback:33
> hugepages-2048kB/stats/swpout:2
> hugepages-2048kB/stats/swpout_fallback:0
> hugepages-1024kB/stats/swpout:0
> hugepages-1024kB/stats/swpout_fallback:0
> hugepages-64kB/stats/swpout:31255
> hugepages-64kB/stats/swpout_fallback:3112
> hugepages-16kB/stats/swpout:29931
> hugepages-16kB/stats/swpout_fallback:3397
> hugepages-256kB/stats/swpout:5223
> hugepages-256kB/stats/swpout_fallback:2351
> hugepages-128kB/stats/swpout:25600
> hugepages-128kB/stats/swpout_fallback:2194
>
> High order (256k) swapout rate are significantly higher, 512k is now
> possible, which indicate high orders are better protected, lower
> orders are sacrificed but seems worth it.

Yes.  I think that this reflects another aspect of the problem.  In some
situations, it's better to steal one high-order cluster and use it for
order-0 allocation instead of scattering order-0 allocation in random
high-order clusters.

--
Best Regards,
Huang, Ying