Le dimanche 02 mai 2010 à 23:54 +0200, Andi Kleen a écrit :

On Sun, May 02, 2010 at 11:45:55PM +0200, Eric Dumazet wrote:

quoted

Tests just prove the reverse.

What do you mean? 

Test I did this week with Jamal.

We first set a "ee" rps mask, because all NIC interrupts were handled by
CPU0, and Jamal thought like you, that not using cpu4 would give better
performance.

But using "fe" mask gave me a bonus, from ~700.000 pps to ~800.000 pps

CPU : E5450  @3.00GHz
Two quad-core cpus in the machine, tg3 NIC.

With RPS, CPU0 does not a lot of things, just talk with the NIC, bring a
few cache lines per packet and dispatch it to a slave cpu.

HT (especially Nehalem HT) is useful for a wide range of workloads.
Just handling network interrupts for its thread sibling is not one of them.

Thats the theory, now in practice I see different results.

Of course, this might be related to hash distribution being different
and more uniform.

I should redo the test with many more flows.