Jakub Kicinski <[email protected]> writes:

> Extend ynltool to compute HW GRO savings metric - how many
> packets has HW GRO been able to save the kernel from seeing.
>
> Note that this definition does not actually take into account
> whether the segments were or weren't eligible for HW GRO.
> If a machine is receiving all-UDP traffic - new metric will show
> HW-GRO savings of 0%. Conversely since the super-packet still
> counts as a received packet, savings of 100% is not achievable.
> Perfect HW-GRO on a machine with 4k MTU and 64kB super-frames
> would show ~93.75% savings. With 1.5k MTU we may see up to
> ~97.8% savings (if my math is right).
>
> Example after 10 sec of iperf on a freshly booted machine
> with 1.5k MTU:
>
>   $ ynltool qstats show
>   eth0     rx-packets:  40681280               rx-bytes:   61575208437
>         rx-alloc-fail:         0      rx-hw-gro-packets:       1225133
>                                  rx-hw-gro-wire-packets:      40656633
>   $ ynltool qstats hw-gro
>   eth0: 96.9% savings
>
> None of the NICs I have access to can report "missed" HW-GRO
> opportunities so computing a true "effectiveness" metric
> is not possible. One could also argue that effectiveness metric
> is inferior in environments where we control both senders and
> receivers, the savings metrics will capture both regressions
> in receiver's HW GRO effectiveness but also regressions in senders
> sending smaller TSO trains. And we care about both. The main
> downside is that it's hard to tell at a glance how well the NIC
> is doing because the savings will be dependent on traffic patterns.
>
> Signed-off-by: Jakub Kicinski <[email protected]>

Reviewed-by: Petr Machata <[email protected]>

Reply via email to