Re: [PATCH RESEND 0/9] sched: cpumask: improve on cpumask_local_spread() locality

[Tariq Toukan]

On 21/01/2023 6:24, Yury Norov wrote:
> cpumask_local_spread() currently checks local node for presence of i'th
> CPU, and then if it finds nothing makes a flat search among all non-local
> CPUs. We can do it better by checking CPUs per NUMA hops.
>
> This has significant performance implications on NUMA machines, for example
> when using NUMA-aware allocated memory together with NUMA-aware IRQ
> affinity hints.
>
> Performance tests from patch 8 of this series for mellanox network
> driver show:
>
>    TCP multi-stream, using 16 iperf3 instances pinned to 16 cores (with aRFS on).
>    Active cores: 64,65,72,73,80,81,88,89,96,97,104,105,112,113,120,121
>    
>    +-------------------------+-----------+------------------+------------------+
>    |                         | BW (Gbps) | TX side CPU util | RX side CPU util |
>    +-------------------------+-----------+------------------+------------------+
>    | Baseline                | 52.3      | 6.4 %            | 17.9 %           |
>    +-------------------------+-----------+------------------+------------------+
>    | Applied on TX side only | 52.6      | 5.2 %            | 18.5 %           |
>    +-------------------------+-----------+------------------+------------------+
>    | Applied on RX side only | 94.9      | 11.9 %           | 27.2 %           |
>    +-------------------------+-----------+------------------+------------------+
>    | Applied on both sides   | 95.1      | 8.4 %            | 27.3 %           |
>    +-------------------------+-----------+------------------+------------------+
>    
>    Bottleneck in RX side is released, reached linerate (~1.8x speedup).
>    ~30% less cpu util on TX.
>
> This series was supposed to be included in v6.2, but that didn't happen. It
> spent enough in -next without any issues, so I hope we'll finally see it
> in v6.3.
>
> I believe, the best way would be moving it with scheduler patches, but I'm
> OK to try again with bitmap branch as well.

Now that Yury dropped several controversial bitmap patches form the PR, 
the rest are mostly in sched, or new API that's used by sched.

Valentin, what do you think? Can you take it to your sched branch?

> Tariq Toukan (1):
>    net/mlx5e: Improve remote NUMA preferences used for the IRQ affinity
>      hints
>
> Valentin Schneider (2):
>    sched/topology: Introduce sched_numa_hop_mask()
>    sched/topology: Introduce for_each_numa_hop_mask()
>
> Yury Norov (6):
>    lib/find: introduce find_nth_and_andnot_bit
>    cpumask: introduce cpumask_nth_and_andnot
>    sched: add sched_numa_find_nth_cpu()
>    cpumask: improve on cpumask_local_spread() locality
>    lib/cpumask: reorganize cpumask_local_spread() logic
>    lib/cpumask: update comment for cpumask_local_spread()
>
>   drivers/net/ethernet/mellanox/mlx5/core/eq.c | 18 +++-
>   include/linux/cpumask.h                      | 20 +++++
>   include/linux/find.h                         | 33 +++++++
>   include/linux/topology.h                     | 33 +++++++
>   kernel/sched/topology.c                      | 90 ++++++++++++++++++++
>   lib/cpumask.c                                | 52 ++++++-----
>   lib/find_bit.c                               |  9 ++
>   7 files changed, 230 insertions(+), 25 deletions(-)