On Sun, 20 Oct 2019 16:07:13 -0500 Zvi Effron <zeffron@xxxxxxxxxxxxx> wrote: > Hello, > > I'm trying to get some additional performance out of XDP on the i40e > driver (under kernel 5.0.7, Fedora 28) by adjusting the queues to be > one per desired core. I have several cores isolated for network > traffic, so I'm adjusting the queues to be one per isolated core (with > IRQ adjustments to match) as opposed to the default of one per online > core. I've noticed that when I adjust the number of queues down, I see > packet loss proportional to the number of queues missing. For example, > if I drop the queues from the default of 36 down to 34, I see ~6% > packet loss. > > Using the trace_pipe, I've managed to confirm that my program is > successfully calling bpf_redirect_map() for all of the packet (no > losst). I found the xdp_monitor tutorial[1] and ran it, and it shows > the loss showing up as XDP_REDIRECT errors (which I believe means the > raw_tracepoint xdp:xdp_redirect_map_err). I usually use standard perf-record: $ perf record -e xdp:xdp_redirect_map_err sleep 1 And then look at detailed results with: $ perf script As this is a standard tracepoint, you can lookup docs for howto use tracepoints. To debug closer you likely want to extract other info, and maybe aggregate. I will recommend looking closer at bpftrace tool[2]. [2] https://github.com/iovisor/bpftrace/blob/master/docs/tutorial_one_liners.md > At this point, I'm a bit stuck on how to proceed further to determine > what is causing the redirect error and where the packet is being > dropped. > > Any help in pointing me in the right direction for debugging > techniques would be greatly appreciated. > > Thanks! > --Zvi > > [1]: https://github.com/xdp-project/xdp-tutorial/blob/master/tracing02-xdp-monitor -- Best regards, Jesper Dangaard Brouer MSc.CS, Principal Kernel Engineer at Red Hat LinkedIn: http://www.linkedin.com/in/brouer
|