Nikita V. Shirokov

Feels like bpf program is not jitted. Could you please run bpftool prog list and bpftool map list

Also in perf report (which was taken with -ag) Please show the output filtered with bpf keyword (/ and than type bpf)

No idea. For some reason bpf program seems slow in bpf code itself. At this point the only idea is to build perf with bpf support (link against that library...

All of that is memory accesses. It feels like there is some issue with it. Either slow memory or system is low on memory. Or tlb is trashed. What the...

Can you run perf to collect counters for "cycles,stalled-cycles-fronted,stalled-cycles-backend". The only explanation that make sense for mov to / From memory to be high on cpu is for some reason...

Yeah. Stalls are way too hard. And they are not even on map accesses. Anyway I think there is some issue with hardware? Do you have any other test machine...

That seems pretty strange. I do not think that it is related to NIC. Seems like some strange memory related hardware issue/ specific. I will post later today how to...

I will try to reply tomorrow

So you would need to build katran's bpf program with INLINE_DECAP define set https://github.com/facebookincubator/katran/blob/main/katran/lib/bpf/balancer_consts.h#L352 Than you would need to run katran_tester with flag perf_testing=true https://github.com/facebookincubator/katran/blob/main/katran/lib/testing/katran_tester.cpp#L57 As example of how to...

12ns per packet are not realistic. You probably forget to set inline_decap define while was building bpf program. Also what are not custom values for defines you have ? Tests...