Willy Tarreau <w...@1wt.eu> wrote: > On Sun, Aug 09, 2020 at 06:30:17PM +0000, George Spelvin wrote: > > Even something simple like buffering 8 TSC samples, and adding them > > at 32-bit offsets across the state every 8th call, would make a huge > > difference. > > Doing testing on real hardware showed that retrieving the TSC on every > call had a non negligible cost, causing a loss of 2.5% on the accept() > rate and 4% on packet rate when using iptables -m statistics. However > I reused your idea of accumulating old TSCs to increase the uncertainty > about their exact value, except that I retrieve it only on 1/8 calls > and use the previous noise in this case. With this I observe the same > performance as plain 5.8. Below are the connection rates accepted on > a single core : > > 5.8 5.8+patch 5.8+patch+tsc > 192900-197900 188800->192200 194500-197500 (conn/s) > > This was on a core i7-8700K. I looked at the asm code for the function > and it remains reasonably light, in the same order of complexity as the > original one, so I think we could go with that. > > My proposed change is below, in case you have any improvements to suggest.
As this relates to networking, you could also hook perturbation into rx/tx softirq processing. E.g. once for each new napi poll round or only once for each softnet invocation, depending on cost. IIRC the proposed draft left a unused prandom_seed() stub around, you could re-use that to place extra data to include in the hash in percpu data.
