lkml.org 
[lkml]   [2023]   [Jul]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [PATCH v2 0/5] variable-order, large folios for anonymous memory
From
On 03.07.23 15:53, Ryan Roberts wrote:
> Hi All,
>
> This is v2 of a series to implement variable order, large folios for anonymous
> memory. The objective of this is to improve performance by allocating larger
> chunks of memory during anonymous page faults. See [1] for background.
>
> I've significantly reworked and simplified the patch set based on comments from
> Yu Zhao (thanks for all your feedback!). I've also renamed the feature to
> VARIABLE_THP, on Yu's advice.
>
> The last patch is for arm64 to explicitly override the default
> arch_wants_pte_order() and is intended as an example. If this series is accepted
> I suggest taking the first 4 patches through the mm tree and the arm64 change
> could be handled through the arm64 tree separately. Neither has any build
> dependency on the other.
>
> The one area where I haven't followed Yu's advice is in the determination of the
> size of folio to use. It was suggested that I have a single preferred large
> order, and if it doesn't fit in the VMA (due to exceeding VMA bounds, or there
> being existing overlapping populated PTEs, etc) then fallback immediately to
> order-0. It turned out that this approach caused a performance regression in the
> Speedometer benchmark. With my v1 patch, there were significant quantities of
> memory which could not be placed in the 64K bucket and were instead being
> allocated for the 32K and 16K buckets. With the proposed simplification, that
> memory ended up using the 4K bucket, so page faults increased by 2.75x compared
> to the v1 patch (although due to the 64K bucket, this number is still a bit
> lower than the baseline). So instead, I continue to calculate a folio order that
> is somewhere between the preferred order and 0. (See below for more details).
>
> The patches are based on top of v6.4 plus Matthew Wilcox's set_ptes() series
> [2], which is a hard dependency. I have a branch at [3].
>
>
> Changes since v1 [1]
> --------------------
>
> - removed changes to arch-dependent vma_alloc_zeroed_movable_folio()
> - replaced with arch-independent alloc_anon_folio()
> - follows THP allocation approach
> - no longer retry with intermediate orders if allocation fails
> - fallback directly to order-0
> - remove folio_add_new_anon_rmap_range() patch
> - instead add its new functionality to folio_add_new_anon_rmap()
> - remove batch-zap pte mappings optimization patch
> - remove enabler folio_remove_rmap_range() patch too
> - These offer real perf improvement so will submit separately
> - simplify Kconfig
> - single FLEXIBLE_THP option, which is independent of arch
> - depends on TRANSPARENT_HUGEPAGE
> - when enabled default to max anon folio size of 64K unless arch
> explicitly overrides
> - simplify changes to do_anonymous_page():
> - no more retry loop
>
>
> Performance
> -----------
>
> Below results show 3 benchmarks; kernel compilation with 8 jobs, kernel
> compilation with 80 jobs, and speedometer 2.0 (a javascript benchmark running in
> Chromium). All cases are running on Ampere Altra with 1 NUMA node enabled,
> Ubuntu 22.04 and XFS filesystem. Each benchmark is repeated 15 times over 5
> reboots and averaged.
>
> 'anonfolio-lkml-v1' is the v1 patchset at [1]. 'anonfolio-lkml-v2' is this v2
> patchset. 'anonfolio-lkml-v2-simple-order' is anonfolio-lkml-v2 but with the
> order selection simplification that Yu Zhao suggested - I'm trying to justify
> here why I did not follow the advice.
>
>
> Kernel compilation with 8 jobs:
>
> | kernel | real-time | kern-time | user-time |
> |:-------------------------------|------------:|------------:|------------:|
> | baseline-4k | 0.0% | 0.0% | 0.0% |
> | anonfolio-lkml-v1 | -5.3% | -42.9% | -0.6% |
> | anonfolio-lkml-v2-simple-order | -4.4% | -36.5% | -0.4% |
> | anonfolio-lkml-v2 | -4.8% | -38.6% | -0.6% |
>
> We can see that the simple-order approach is responsible for a regression of
> 0.4%.
>
>
> Kernel compilation with 80 jobs:
>
> | kernel | real-time | kern-time | user-time |
> |:-------------------------------|------------:|------------:|------------:|
> | baseline-4k | 0.0% | 0.0% | 0.0% |
> | anonfolio-lkml-v1 | -4.6% | -45.7% | 1.4% |
> | anonfolio-lkml-v2-simple-order | -4.7% | -40.2% | -0.1% |
> | anonfolio-lkml-v2 | -5.0% | -42.6% | -0.3% |
>
> simple-order costs 0.3 % here. v2 is actually performing higher than v1 due to
> fixing the v1 regression on user-time.
>
>
> Speedometer 2.0:
>
> | kernel | runs_per_min |
> |:-------------------------------|---------------:|
> | baseline-4k | 0.0% |
> | anonfolio-lkml-v1 | 0.7% |
> | anonfolio-lkml-v2-simple-order | -0.9% |
> | anonfolio-lkml-v2 | 0.5% |
>
> simple-order regresses performance by 0.9% vs the baseline, for a total negative
> swing of 1.6% vs v1. This is fixed by keeping the more complex order selection
> mechanism from v1.
>
>
> The remaining (kernel time) performance gap between v1 and v2 for the above
> benchmarks is due to the removal of the "batch zap" patch in v2. Adding that
> back in gives us the performance back. I intend to submit that as a separate
> series once this series is accepted.
>
>
> [1] https://lore.kernel.org/linux-mm/20230626171430.3167004-1-ryan.roberts@arm.com/
> [2] https://lore.kernel.org/linux-mm/20230315051444.3229621-1-willy@infradead.org/
> [3] https://gitlab.arm.com/linux-arm/linux-rr/-/tree/features/granule_perf/anonfolio-lkml_v2
>
> Thanks,
> Ryan

Hi Ryan,

is page migration already working as expected (what about page
compaction?), and do we handle migration -ENOMEM when allocating a
target page: do we split an fallback to 4k page migration?

--
Cheers,

David / dhildenb

\
 
 \ /
  Last update: 2023-07-05 21:40    [W:0.419 / U:0.172 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site