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

From: David Hildenbrand
Date: Wed Jan 31 2024 - 07:56:43 EST

On 31.01.24 13:37, Ryan Roberts wrote:
On 31/01/2024 11:49, Ryan Roberts wrote:
On 31/01/2024 11:28, David Hildenbrand wrote:
On 31.01.24 12:16, Ryan Roberts wrote:
On 31/01/2024 11:06, David Hildenbrand wrote:
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.

Numbers for v1 of the series, both on top of 6.8-rc1 and rebased to the same
mm-unstable base as v3 of the series (first 2 rows are from what I just posted
for context):

| kernel | mean_rel | std_rel |
|:-------------------|-----------:|----------:|
| mm-unstabe (base) | 0.0% | 1.1% |
| mm-unstable + v3 | 16.7% | 0.8% |
| mm-unstable + v1 | -2.5% | 1.7% |
| v6.8-rc1 + v1 | -6.6% | 1.1% |

So all looks good with v1. And seems to suggest mm-unstable has regressed by ~4%
vs v6.8-rc1. Is this really a useful benchmark? Does the raw performance of
fork() syscall really matter? Evidence suggests its moving all over the place -
breath on the code and it changes - not a great place to be when using the test
for gating purposes!

Still with the old tests - I'll move to the new ones now.




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?

Nope; running the exact same kernel binary and user space on Altra, I see
sensible numbers;

fork order-0: -1.3%
fork order-9: -7.6%
dontneed order-0: -0.5%
dontneed order-9: 0.1%
munmap order-0: 0.0%
munmap order-9: -67.9%

So I guess some pipelining issue that causes the M2 to stall more?

With one effective added folio_test_large(), it could only be a code layout
problem? Or the compiler does something stupid, but you say that you run the
exact same kernel binary, so that doesn't make sense.

Yup, same binary. We know this code is very sensitive - 1 cycle makes a big
difference. So could easily be code layout, branch prediction, etc...


I'm also surprised about the dontneed vs. munmap numbers.

You mean the ones for Altra that I posted? (I didn't post any for M2). The altra
numbers look ok to me; dontneed has no change, and munmap has no change for
order-0 and is massively improved for order-9.

Doesn't make any sense
(again, there was this VMA merging problem but it would still allow for batching
within a single VMA that spans exactly one large folio).

What are you using as baseline? Really just mm-unstable vs. mm-unstable+patches?

yes. except for "v6.8-rc1 + v1" above.


Let's see if the new test changes the numbers you measure.

Nope: looks the same. I've taken my test harness out of the picture and done
everything manually from the ground up, with the old tests and the new. Headline
is that I see similar numbers from both.

I took me a while to get really reproducible numbers on Intel. Most importantly:
* Set a fixed CPU frequency, disabling any boost and avoiding any
thermal throttling.
* Pin the test to CPUs and set a nice level.

Another thing is, to avoid systems where you can have NUMA effects within a single socket. Otherwise, memory access latency is just random and depends on what the buddy enjoys giving you.

But you seem to get the same +17 even after reboots, so that indicates that the CPU is not happy about the code for some reason. And the weird thing is, that nothing significantly changed for order-0 folios between v1 and v3 that could explain any of this.

I'm not worried about 5% or so, nobody cares. But it would be good to have at least an explanation why only that system shows +17%.


Some details:
- I'm running for 10 seconds then averaging the output

Same here.

- test is bimodal; first run (of 10 seconds) after boot is a bit faster on
average (up to 10%) than the rest; I could guess this is due to the memory
being allocated more contiguously the first few times through, so struct
pages have better locality, but that's a guess.

I think it also has to do with the PCP lists, and the high-pcp auto tuning (I played with disabling that). Running on a freshly booted system gave me reproducible results.

But yes: I was observing something similar on AMD EPYC, where you get consecutive pages from the buddy, but once you allocate from the PCP it might no longer be consecutive.

- test is 5-10% slower when output is printed to terminal vs when redirected to
file. I've always effectively been redirecting. Not sure if this overhead
could start to dominate the regression and that's why you don't see it?

That's weird, because we don't print while measuring? Anyhow, 5/10% variance on some system is not the end of the world.


I'm inclined to run this test for the last N kernel releases and if the number
moves around significantly, conclude that these tests don't really matter.
Otherwise its an exercise in randomly refactoring code until it works well, but
that's just overfitting to the compiler and hw. What do you think?

Personally, I wouldn't lose sleep if you see weird, unexplainable behavior on some system (not even architecture!). Trying to optimize for that would indeed be random refactorings.

But I would not be so fast to say that "these tests don't really matter" and then go wild and degrade them as much as you want. There are use cases that care about fork performance especially with order-0 pages -- such as Redis.

--
Cheers,

David / dhildenb