Re: [PATCH v4 0/6] Swap-out mTHP without splitting

[Ryan Roberts <ryan.roberts@arm.com>]

On 13/03/2024 01:15, Huang, Ying wrote:
> Ryan Roberts <ryan.roberts@arm.com> writes:
> 
>> On 12/03/2024 08:49, Ryan Roberts wrote:
>>> On 12/03/2024 08:01, Huang, Ying wrote:
>>>> Ryan Roberts <ryan.roberts@arm.com> writes:
>>>>
>>>>> Hi All,
>>>>>
>>>>> This series adds support for swapping out multi-size THP (mTHP) without needing
>>>>> to first split the large folio via split_huge_page_to_list_to_order(). It
>>>>> closely follows the approach already used to swap-out PMD-sized THP.
>>>>>
>>>>> There are a couple of reasons for swapping out mTHP without splitting:
>>>>>
>>>>>   - Performance: It is expensive to split a large folio and under extreme memory
>>>>>     pressure some workloads regressed performance when using 64K mTHP vs 4K
>>>>>     small folios because of this extra cost in the swap-out path. This series
>>>>>     not only eliminates the regression but makes it faster to swap out 64K mTHP
>>>>>     vs 4K small folios.
>>>>>
>>>>>   - Memory fragmentation avoidance: If we can avoid splitting a large folio
>>>>>     memory is less likely to become fragmented, making it easier to re-allocate
>>>>>     a large folio in future.
>>>>>
>>>>>   - Performance: Enables a separate series [4] to swap-in whole mTHPs, which
>>>>>     means we won't lose the TLB-efficiency benefits of mTHP once the memory has
>>>>>     been through a swap cycle.
>>>>>
>>>>> I've done what I thought was the smallest change possible, and as a result, this
>>>>> approach is only employed when the swap is backed by a non-rotating block device
>>>>> (just as PMD-sized THP is supported today). Discussion against the RFC concluded
>>>>> that this is sufficient.
>>>>>
>>>>>
>>>>> Performance Testing
>>>>> ===================
>>>>>
>>>>> I've run some swap performance tests on Ampere Altra VM (arm64) with 8 CPUs. The
>>>>> VM is set up with a 35G block ram device as the swap device and the test is run
>>>>> from inside a memcg limited to 40G memory. I've then run `usemem` from
>>>>> vm-scalability with 70 processes, each allocating and writing 1G of memory. I've
>>>>> repeated everything 6 times and taken the mean performance improvement relative
>>>>> to 4K page baseline:
>>>>>
>>>>> | alloc size |            baseline |       + this series |
>>>>> |            |  v6.6-rc4+anonfolio |                     |
>>>>> |:-----------|--------------------:|--------------------:|
>>>>> | 4K Page    |                0.0% |                1.4% |
>>>>> | 64K THP    |              -14.6% |               44.2% |
>>>>> | 2M THP     |               87.4% |               97.7% |
>>>>>
>>>>> So with this change, the 64K swap performance goes from a 15% regression to a
>>>>> 44% improvement. 4K and 2M swap improves slightly too.
>>>>
>>>> I don't understand why the performance of 2M THP improves.  The swap
>>>> entry allocation becomes a little slower.  Can you provide some
>>>> perf-profile to root cause it?
>>>
>>> I didn't post the stdev, which is quite large (~10%), so that may explain some
>>> of it:
>>>
>>> | kernel   |   mean_rel |   std_rel |
>>> |:---------|-----------:|----------:|
>>> | base-4K  |       0.0% |      5.5% |
>>> | base-64K |     -14.6% |      3.8% |
>>> | base-2M  |      87.4% |     10.6% |
>>> | v4-4K    |       1.4% |      3.7% |
>>> | v4-64K   |      44.2% |     11.8% |
>>> | v4-2M    |      97.7% |     13.3% |
>>>
>>> Regardless, I'll do some perf profiling and post results shortly.
>>
>> I did a lot more runs (24 for each config) and meaned them to try to remove the
>> noise in the measurements. It's now only showing a 4% improvement for 2M. So I
>> don't think the 2M improvement is real:
>>
>> | kernel   |   mean_rel |   std_rel |
>> |:---------|-----------:|----------:|
>> | base-4K  |       0.0% |      3.2% |
>> | base-64K |      -9.1% |     10.1% |
>> | base-2M  |      88.9% |      6.8% |
>> | v4-4K    |       0.5% |      3.1% |
>> | v4-64K   |      44.7% |      8.3% |
>> | v4-2M    |      93.3% |      7.8% |
>>
>> Looking at the perf data, the only thing that sticks out is that a big chunk of
>> time is spent in during contpte_convert(), called as a result of
>> try_to_unmap_one(). This is present in both the before and after configs.
>>
>> This is an arm64 function to "unfold" contpte mappings. Essentially, the PMD is
>> being split during shrink_folio_list()  with TTU_SPLIT_HUGE_PMD, meaning the
>> THPs are PTE-mapped in contpte blocks. Then we are unmapping each pte one-by-one
>> which means the contpte block needs to be unfolded. I think try_to_unmap_one()
>> could potentially be optimized to batch unmap a contiguously mapped folio and
>> avoid this unfold. But that would be an independent and separate piece of work.
> 
> Thanks for more data and detailed explanation.

And thanks for your review! I'll address all your comments (and any others that
I get in the meantime) and repost after the merge window. It would be great if
we can get this in for v6.10.

> 
> --
> Best Regards,
> Huang, Ying