Sebastien Roy wrote:
> Garrett D'Amore wrote:
>> I'd bet, with a high degree of confidence, that if you tried doing an
>> IP forwarding test with small packets, you'd find that performance
>> improvements are _not_ in the noise, unless your noise filter is set
>> too low.
>
> I used the DIY perf PIT, and my observations are that there is a lot
> of noise in their test runs due to small sample sizes.
You can actually get very, very reproducible performance results without
needing large sample sizes. Within a test that takes just 10 minutes to
run using the Smartbits hardware, I can get quite repeatable results for
IP forwarding.
Of course, I've been very very focused on the most extreme end of CPU
bound processing, which is trying to route 64-byte packets...
>
>> Another way to test would be to try doing performance runs with a 10g
>> card using something _other_ than TCP (UDP rx would be good). Look
>> at the improvements to the packets-per-second count rather than the
>> thruput numbers. :-)
>>
>> If you're not CPU bound, you won't see the benefit.
>
> Right. I wanted to run sanity tests to verify that I had not made
> things worse. Performance improvement was not a primary goal, so I
> did not spend many resources evaluating potential positive performance
> impact.
Sure. Once your changes get into the gate, I'd be interested to see
what happens to our Sitara numbers. I had already #if 0'd at least the
mobile IP checks in front of ip_fast_forward(), in anticipation of these
changes. ;-)
-- Garrett
>
> -Seb
