Re: [PATCH] Load balancing problem in 2.6.2-mm1

[Nick Piggin]

Rick Lindsley wrote:

>    Yep. I've argued for fairness here, and that is presently what
>    we get. Between nodes the threshold should probably be higher
>    though.
>
> While I like the idea of a self-tuning scheduler, when combined with
> this new sched_domain algorithm it's hard to tell if the tuning or the
> algorithm is at fault when we get results we don't like.  Have you done
> much running with the auto-tuning turned off, using the old values,
> to see the impact (positive or negative) that just the new algorithm has?

I'm not sure what you mean by self-tuning. Do you mean the scheduling
backoff stuff? Because that makes very little difference on a 16-way
NUMAQ. However it becomes critical for SGI above around 128 CPUs IIRC
so I just kept it in mind when doing sched domains.

The new balancing calculations are definitely a win in my tests. One
tiny regression (the order of 1%) I saw on the NUMAQ was tbench due to
increased idle time. But I'll still take it as a win because we were
doing nearly 1000 times less inter node balancing.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/