Re: [PATCH v3 00/15] mm/memory: optimize fork() with PTE-mapped THP

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 31.01.24 11:43, Ryan Roberts wrote:
On 29/01/2024 12:46, David Hildenbrand wrote:
Now that the rmap overhaul[1] is upstream that provides a clean interface
for rmap batching, let's implement PTE batching during fork when processing
PTE-mapped THPs.

This series is partially based on Ryan's previous work[2] to implement
cont-pte support on arm64, but its a complete rewrite based on [1] to
optimize all architectures independent of any such PTE bits, and to
use the new rmap batching functions that simplify the code and prepare
for further rmap accounting changes.

We collect consecutive PTEs that map consecutive pages of the same large
folio, making sure that the other PTE bits are compatible, and (a) adjust
the refcount only once per batch, (b) call rmap handling functions only
once per batch and (c) perform batch PTE setting/updates.

While this series should be beneficial for adding cont-pte support on
ARM64[2], it's one of the requirements for maintaining a total mapcount[3]
for large folios with minimal added overhead and further changes[4] that
build up on top of the total mapcount.

Independent of all that, this series results in a speedup during fork with
PTE-mapped THP, which is the default with THPs that are smaller than a PMD
(for example, 16KiB to 1024KiB mTHPs for anonymous memory[5]).

On an Intel Xeon Silver 4210R CPU, fork'ing with 1GiB of PTE-mapped folios
of the same size (stddev < 1%) results in the following runtimes
for fork() (shorter is better):

Folio Size | v6.8-rc1 |      New | Change
------------------------------------------
       4KiB | 0.014328 | 0.014035 |   - 2%
      16KiB | 0.014263 | 0.01196  |   -16%
      32KiB | 0.014334 | 0.01094  |   -24%
      64KiB | 0.014046 | 0.010444 |   -26%
     128KiB | 0.014011 | 0.010063 |   -28%
     256KiB | 0.013993 | 0.009938 |   -29%
     512KiB | 0.013983 | 0.00985  |   -30%
    1024KiB | 0.013986 | 0.00982  |   -30%
    2048KiB | 0.014305 | 0.010076 |   -30%

Just a heads up that I'm seeing some strange results on Apple M2. Fork for
order-0 is seemingly costing ~17% more. I'm using GCC 13.2 and was pretty sure I
didn't see this problem with version 1; although that was on a different
baseline and I've thrown the numbers away so will rerun and try to debug this.


So far, on my x86 tests (Intel, AMD EPYC), I was not able to observe this. fork() for order-0 was consistently effectively unchanged. Do you observe that on other ARM systems as well?


| kernel      |   mean_rel |   std_rel |
|:------------|-----------:|----------:|
| mm-unstable |       0.0% |      1.1% |
| patch 1     |      -2.3% |      1.3% |
| patch 10    |      -2.9% |      2.7% |
| patch 11    |      13.5% |      0.5% |
| patch 12    |      15.2% |      1.2% |
| patch 13    |      18.2% |      0.7% |
| patch 14    |      20.5% |      1.0% |
| patch 15    |      17.1% |      1.6% |
| patch 15    |      16.7% |      0.8% |

fork for order-9 is looking good (-20%), and for the zap series, munmap is
looking good, but dontneed is looking poor for both order-0 and 9. But one thing
at a time... let's concentrate on fork order-0 first.

munmap and dontneed end up calling the exact same call paths. So a big performance difference is rather surprising and might indicate something else.

(I think I told you that I was running in some kind of VMA merging problem where one would suddenly get with my benchmark 1 VMA per page. The new benchmark below works around that, but I am not sure if that was fixed in the meantime)

VMA merging can of course explain a big difference in fork and munmap vs. dontneed times, especially when comparing different code base where that VMA merging behavior was different.


Note that I'm still using the "old" benchmark code. Could you resend me the link
to the new code? Although I don't think there should be any effect for order-0
anyway, if I understood your changes correctly?

This is the combined one (small and large PTEs):

https://gitlab.com/davidhildenbrand/scratchspace/-/raw/main/pte-mapped-folio-benchmarks.c?inline=false

--
Cheers,

David / dhildenb





[Index of Archives]     [Kernel Development]     [DCCP]     [Linux ARM Development]     [Linux]     [Photo]     [Yosemite Help]     [Linux ARM Kernel]     [Linux SCSI]     [Linux x86_64]     [Linux Hams]

  Powered by Linux