Re: [Regression] mqueue performance degradation after "The new cgroup slab memory controller" patchset.

[Roman Gushchin <roman.gushchin@linux.dev>]

On Mon, Dec 05, 2022 at 02:55:48PM +0000, Luther, Sven wrote:
> #regzbot ^introduced 10befea91b61c4e2c2d1df06a2e978d182fcf792
> 
> We are making heavy use of mqueues, and noticed a degradation of performance between 4.18 & 5.10 linux kernels.
> 
> After a gross per-version tracing, we did kernel bisection between 5.8 and 5.9
> and traced the issue to a 10 patches (of which 9 where skipped as they didn't boot) between:
> 
> 
> commit 10befea91b61c4e2c2d1df06a2e978d182fcf792 (HEAD, refs/bisect/bad)
> Author: Roman Gushchin <guro@fb.com>
> Date:   Thu Aug 6 23:21:27 2020 -0700
> 
>     mm: memcg/slab: use a single set of kmem_caches for all allocations
> 
> and:
> 
> commit 286e04b8ed7a04279ae277f0f024430246ea5eec (refs/bisect/good-286e04b8ed7a04279ae277f0f024430246ea5eec)
> Author: Roman Gushchin <guro@fb.com>
> Date:   Thu Aug 6 23:20:52 2020 -0700
> 
>     mm: memcg/slab: allocate obj_cgroups for non-root slab pages
> 
> All of them are part of the "The new cgroup slab memory controller" patchset:
> 
>   https://lore.kernel.org/all/20200623174037.3951353-18-guro@fb.com/T/
> 
> from Roman Gushchin, which moves the accounting for page level to the object level.
> 
> Measurements where done using the a test programmtest, which measures mix/average/max time mqueue_send/mqueue_rcv,
> and average for getppid, both measured over 100 000 runs. Results are shown in the following table
> 
> +----------+--------------------------+-------------------------+----------------+
> | kernel   |    mqueue_rcv (ns)       | mqueue_send (ns)        |    getppid     |
> | version  | min avg  max   variation | min avg max   variation | (ns) variation |
> +----------+--------------------------+-------------------------+----------------+
> | 4.18.45  | 351 382 17533     base   | 383 410 13178     base  | 149      base  |
> | 5.8-good | 380 392  7156   -2,55%   | 376 384  6225    6,77%  | 169   -11,83%  |
> | 5.8-bad  | 524 530  5310  -27,92%   | 512 519  8775  -21,00%  | 169   -11,83%  |
> | 5.10     | 520 533  4078  -28,33%   | 518 534  8108  -23,22%  | 167   -10,78%  |
> | 5.15     | 431 444  8440  -13,96%   | 425 437  6170   -6,18%  | 171   -12,87%  |
> | 6.03     | 474 614  3881  -37,79%   | 482 693   931  -40,84%  | 171   -12,87%  |
> +----------+--------------------------+-------------------------+-----------------

Hi Sven!

Thank you for the report! As Waiman said, it's not a secret that per-object tracking
makes individual allocations slower, but for the majority of workloads it's well
compensated by significant memory savings and a lower fragmentation.

It seems there is another regression between 5.15 and 6.03, which is a separate
topic, but how big is the real regression between 4.18 and 5.15? The benchmark
shows about 14%, but is you real workload suffering at the same level?
If the answer is yes, the right thing to do is to introduce some sort of
mqueue-specific caching for allocated objects.

Thanks!

Roman