On Tue, Apr 20, 2010 at 1:45 PM, Eric Dumazet <eric.dumazet@gmail.com> wrote:
quoted text
> Le lundi 19 avril 2010 à 22:15 -0700, Tom Herbert a écrit :
>> On Mon, Apr 19, 2010 at 9:08 PM, Changli Gao <xiaosuo@gmail.com> wrote:
>> > rps: send IPIs ASAP
>> >
>> > In order to reduce latency, we'd better send IPIs ASAP to schedule the
>> > corresponding NAPIs.
>> >
>> A design point of RPS is that we generate at most one IPI per CPU per
>> device interrupt, which at least offers some predictable coalescing.
>> With your changes, we would get at most one IPI per packet-- that
>> could represent a lot more of them.  Did you test this to see what the
>> impact is in this regard?
>>
>
> I agree with you Tom. Coalescing IPI is probably better.
>
> If the receiver CPU got a single packet in its RX handling, latency will
> be the same anyway.

I did the "ping -f" test again, and found that the differences of RTT
I got before were noises. It seems your "shortcut net_rps_action()"
patch eliminates the differences.

quoted text
>
> If the receiver CPU got many packets, chance is high we are in a stress
> situation, and coalescing is a win in this case.
>
> I am currently testing a patch to call net_rps_action() at the beginning
> of process_backlog() (if we have a non null ipi_rps_list pointer)
>
> Will post a patch with bench results
>

It sounds like a better idea.


-- 
Regards，
Changli Gao(xiaosuo@gmail.com)
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html