On Fri, 22 Sep 2023, Peter Zijlstra wrote:
Hi! Updated version of patch 15/15 and a few extra patches for testing the FUTEX2_NUMA bits. The last patch (17/15) should never be applied for anything you care about and exists purely because I'm too lazy to generate actual hash-bucket contention. On my 2 node IVB-EP: $ echo FUTEX_SQUASH > /debug/sched/features Effectively reducing each node to 1 bucket. $ numactl -m0 -N0 ./futex_numa -c10 -t2 -n0 -N0 & numactl -m1 -N1 ./futex_numa -c10 -t2 -n0 -N0 ... contenders: 16154935 contenders: 16202472 $ numactl -m0 -N0 ./futex_numa -c10 -t2 -n0 -N0 & numactl -m1 -N1 ./futex_numa -c10 -t2 -n0 -N1 contenders: 48584991 contenders: 48680560 (loop counts, higher is better) Clearly showing how separating the hashes works. The first one runs 10 contenders on each node but forces the (numa) futex to hash to node 0 for both. This ensures all 20 contenders hash to the same bucket and *ouch*. The second one does the same, except now fully separates the nodes. Performance is much improved. Proving the per-node hashing actually works as advertised.
Very nice.
