Re: [PATCH 4/4] sched: bias to target cpu load to reduce task moving

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Jan 07, 2014 at 03:16:32PM +0000, Morten Rasmussen wrote:
> From a load perspective wouldn't it be better to pick the least loaded
> cpu in the group? It is not cheap to implement, but in theory it should
> give less balancing within the group later an less unfairness until it
> happens.

I tried that; see 04f733b4afac5dc93ae9b0a8703c60b87def491e for why it
doesn't work.

> Rotating the cpu is probably good enough for most cases and certainly
> easier to implement.

Indeed.

> The bias continues after they first round of load balance by the other
> cpus?

The cost, yes. Even when perfectly balanced, we still get to iterate the
entire machine computing s[gd]_lb_stats to find out we're good and don't
need to move tasks about.

> Pulling everything to one cpu is not ideal from a performance point of
> view. You loose some available cpu cycles until the balance settles.
> However, it is not easy to do better and maintain scalability at the
> same time.

Right, its part of the cost we pay for scaling better. And rotating this
cost around a bit would alleviate the obvious bias.