Re: [PATCH -mm -v9 2/3] mm, THP, swap: Check whether THP can be split firstly

[Huang\, Ying]

Johannes Weiner <hannes@xxxxxxxxxxx> writes:

> On Wed, Apr 19, 2017 at 03:06:24PM +0800, Huang, Ying wrote:
>> From: Huang Ying <ying.huang@xxxxxxxxx>
>> 
>> To swap out THP (Transparent Huage Page), before splitting the THP,
>> the swap cluster will be allocated and the THP will be added into the
>> swap cache.  But it is possible that the THP cannot be split, so that
>> we must delete the THP from the swap cache and free the swap cluster.
>> To avoid that, in this patch, whether the THP can be split is checked
>> firstly.  The check can only be done racy, but it is good enough for
>> most cases.
>> 
>> With the patchset, the swap out throughput improves 3.6% (from about
>> 4.16GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
>> with 8 processes.  The test is done on a Xeon E5 v3 system.  The swap
>> device used is a RAM simulated PMEM (persistent memory) device.  To
>> test the sequential swapping out, the test case creates 8 processes,
>> which sequentially allocate and write to the anonymous pages until the
>> RAM and part of the swap device is used up.
>> 
>> Cc: Johannes Weiner <hannes@xxxxxxxxxxx>
>> Signed-off-by: "Huang, Ying" <ying.huang@xxxxxxxxx>
>> Acked-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx> [for can_split_huge_page()]
>
> How often does this actually happen in practice? Because all that this
> protects us from is trying to allocate a swap cluster - which with the
> si->free_clusters list really isn't all that expensive - and return it
> again. Unless this happens all the time in practice, this optimization
> seems misplaced.

In addition to allocate/free swap cluster, add/delete to/from the swap
cache will be called too.

> It's especially a little strange because in the other email I asked
> about the need for unlikely() annotations, yet this patch is adding
> branches and checks for what seems to be an unlikely condition into
> the THP hot path.
>
> I'd suggest you drop both these optimization attempts unless there is
> real data proving that they have a measurable impact.

To my surprise too, I found this patch has measurable impact in my
test.  The swap out throughput improves 3.6% in the vm-scalability
swap-w-seq test case with 8 processes.  Details are in the original
patch description.

Best Regards,
Huang, Ying