On Thu, Dec 08, 2022 at 10:22:22PM +0200, Tariq Toukan wrote: > > > On 12/8/2022 8:30 PM, 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 series is inspired by Tariq Toukan and Valentin Schneider's > > "net/mlx5e: Improve remote NUMA preferences used for the IRQ affinity > > hints" > > > > https://patchwork.kernel.org/project/netdevbpf/patch/20220728191203.4055-3-tariqt@xxxxxxxxxx/ > > > > According to their measurements, for mlx5e: > > > > Bottleneck in RX side is released, reached linerate (~1.8x speedup). > > ~30% less cpu util on TX. > > > > This patch makes cpumask_local_spread() traversing CPUs based on NUMA > > distance, just as well, and I expect comparable improvement for its > > users, as in case of mlx5e. > > > > I tested new behavior on my VM with the following NUMA configuration: > > > > root@debian:~# numactl -H > > available: 4 nodes (0-3) > > node 0 cpus: 0 1 2 3 > > node 0 size: 3869 MB > > node 0 free: 3740 MB > > node 1 cpus: 4 5 > > node 1 size: 1969 MB > > node 1 free: 1937 MB > > node 2 cpus: 6 7 > > node 2 size: 1967 MB > > node 2 free: 1873 MB > > node 3 cpus: 8 9 10 11 12 13 14 15 > > node 3 size: 7842 MB > > node 3 free: 7723 MB > > node distances: > > node 0 1 2 3 > > 0: 10 50 30 70 > > 1: 50 10 70 30 > > 2: 30 70 10 50 > > 3: 70 30 50 10 > > > > And the cpumask_local_spread() for each node and offset traversing looks > > like this: > > > > node 0: 0 1 2 3 6 7 4 5 8 9 10 11 12 13 14 15 > > node 1: 4 5 8 9 10 11 12 13 14 15 0 1 2 3 6 7 > > node 2: 6 7 0 1 2 3 8 9 10 11 12 13 14 15 4 5 > > node 3: 8 9 10 11 12 13 14 15 4 5 6 7 0 1 2 3 > > > > v1: https://lore.kernel.org/lkml/20221111040027.621646-5-yury.norov@xxxxxxxxx/T/ > > v2: https://lore.kernel.org/all/20221112190946.728270-3-yury.norov@xxxxxxxxx/T/ > > v3: > > - fix typo in find_nth_and_andnot_bit(); > > - add 5th patch that simplifies cpumask_local_spread(); > > - address various coding style nits. > > > > Yury Norov (5): > > 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 > > > > include/linux/cpumask.h | 20 ++++++++++++++ > > include/linux/find.h | 33 +++++++++++++++++++++++ > > include/linux/topology.h | 8 ++++++ > > kernel/sched/topology.c | 57 ++++++++++++++++++++++++++++++++++++++++ > > lib/cpumask.c | 26 +++++------------- > > lib/find_bit.c | 9 +++++++ > > 6 files changed, 134 insertions(+), 19 deletions(-) > > > > Acked-by: Tariq Toukan <tariqt@xxxxxxxxxx> Thanks Tariq, Jacob and Peter for review. I'll add the series in bitmap-for-next for testing. Still, I think that sched/numa branches would be more suitable. Thanks, Yury
