Re: + sched-use-tasklet-to-call-balancing.patch added to -mm tree

[Ingo Molnar <mingo@elte.hu>]

* Christoph Lameter <clameter@sgi.com> wrote:

> Tasklets are scheduled on the same cpu that triggered the tasklet. 
> They are just moved to other processors if the processor goes down. So 
> that aspect is fine. We just need a tasklet struct per cpu.
> 
> User a per cpu tasklet to schedule rebalancing
> 
> Turns out that tasklets have a flag that only allows one instance to 
> run on all processors. So we need a tasklet structure for each 
> processor.

Per-CPU tasklets are equivalent to softirqs, with extra complexity and 
overhead ontop of it :-)

so please just introduce a rebalance softirq and attach the scheduling 
rebalance tick to it. But i'd suggest to re-test on the 4096-CPU box, 
maybe what 'fixed' your workload was the global serialization of the 
tasklet. With a per-CPU softirq approach we are i think back to the same 
situation that broke your system before.

	Ingo