On Tue, 10 Jul 2018 16:57:14 +0200 "Jason A. Donenfeld" <[email protected]> wrote:
> The latest snapshot will still have the same preemption relaxation > with simd_relax(), but gets performance gains by moving to napi, so > it's still faster overall. If you want the simd_relax() to not take a > hit and get maximum throughput, the right way of doing this is > actually to just disable preemption in your kernel with > PREEMPT_NONE=y. I build a single kernel to use across a diverse park of machines, including servers, routers -- and a few GUI desktops. It is not an option for me to disable preemption entirely in that kernel. (And it would be a hassle to build two or more kernels each time). However those of my hosts which are routers with WG, are NOT the same hosts which are interactive desktops. So I don't want any sacrifices towards interactivity *in WG*. I'll probably test again without simd_relax, but from the past mailing list discussion it seemed like that one doesn't affect things much. In any case, it's great that you found a way to keep performance and increase interactivity at the same time with NAPI. -- With respect, Roman _______________________________________________ WireGuard mailing list [email protected] https://lists.zx2c4.com/mailman/listinfo/wireguard
