On Fri, 2023-05-26 at 08:38 -0700, Haiyang Zhang wrote: > The apc->eth_stats.rx_cqes is one per NIC (vport), and it's on the > frequent and parallel code path of all queues. So, r/w into this > single shared variable by many threads on different CPUs creates a > lot caching and memory overhead, hence perf regression. And, it's > not accurate due to the high volume concurrent r/w. > > For example, a workload is iperf with 128 threads, and with RPS > enabled. We saw perf regression of 25% with the previous patch > adding the counters. And this patch eliminates the regression. > > Since the error path of mana_poll_rx_cq() already has warnings, so > keeping the counter and convert it to a per-queue variable is not > necessary. So, just remove this counter from this high frequency > code path. > > Also, remove the tx_cqes counter for the same reason. We have > warnings & other counters for errors on that path, and don't need > to count every normal cqe processing. FTR, if in future you will need the above counters again, you could re- add them using per-cpu variables to avoid re-introducing the regression addressed here. Cheers, Paolo
