Hangbin Liu <liuhangbin@xxxxxxxxx> writes: > On Thu, Jun 04, 2020 at 02:37:23PM +0200, Toke Høiland-Jørgensen wrote: >> > Now I use the ethtool_stats.pl to count forwarding speed and here is the result: >> > >> > With kernel 5.7(ingress i40e, egress i40e) >> > XDP: >> > bridge: 1.8M PPS >> > xdp_redirect_map: >> > generic mode: 1.9M PPS >> > driver mode: 10.4M PPS >> >> Ah, now we're getting somewhere! :) >> >> > Kernel 5.7 + my patch(ingress i40e, egress i40e) >> > bridge: 1.8M >> > xdp_redirect_map: >> > generic mode: 1.86M PPS >> > driver mode: 10.17M PPS >> >> Right, so this corresponds to a ~2ns overhead (10**9/10400000 - >> 10**9/10170000). This is not too far from being in the noise, I suppose; >> is the difference consistent? > > Sorry, I didn't get, what different consistent do you mean? I meant, how much do the numbers vary between each test run? >> > xdp_redirect_map_multi: >> > generic mode: 1.53M PPS >> > driver mode: 7.22M PPS >> > >> > Kernel 5.7 + my patch(ingress i40e, egress veth) >> > xdp_redirect_map: >> > generic mode: 1.38M PPS >> > driver mode: 4.15M PPS >> > xdp_redirect_map_multi: >> > generic mode: 1.13M PPS >> > driver mode: 3.55M PPS > > With XDP_DROP in veth perr, the number looks much better > > xdp_redirect_map: > generic mode: 1.64M PPS > driver mode: 13.3M PPS > xdp_redirect_map_multi: > generic mode: 1.29M PPS > driver mode: 8.5M PPS Is this for a single interface in both cases? Look a bit odd that you get such a big difference all of a sudden; is the redirect failing in one of those cases (should be a hint in the ethtool stats, I think, otherwise check xdp_monitor)? >> > Kernel 5.7 + my patch(ingress i40e, egress i40e + veth) >> > xdp_redirect_map_multi: >> > generic mode: 1.13M PPS >> > driver mode: 3.47M PPS > > But I don't know why this one get even a little slower.. > > xdp_redirect_map_multi: > generic mode: 0.96M PPS > driver mode: 3.14M PPS Yeah, this does seem a bit odd. Don't have any good ideas off the top of my head, but maybe worth double-checking where the time is spent. You can use 'perf' for this, but you need to make sure it's recording the CPU that is processing packets... -Toke
