Re: [PATCH] sched: wakeup buddy

[Ingo Molnar <mingo@kernel.org>]

* Michael Wang <wangyun@linux.vnet.ibm.com> wrote:

> On 03/11/2013 05:40 PM, Ingo Molnar wrote:
> > 
> > * Michael Wang <wangyun@linux.vnet.ibm.com> wrote:
> > 
> >> Hi, Ingo
> >>
> >> On 03/11/2013 04:21 PM, Ingo Molnar wrote:
> >> [snip]
> >>>
> >>> I have actually written the prctl() approach before, for instrumentation 
> >>> purposes, and it does wonders to system analysis.
> >>
> >> The idea sounds great, we could get many new info to implement more
> >> smart scheduler, that's amazing :)
> >>
> >>>
> >>> Any objections?
> >>
> >> Just one concern, may be I have misunderstand you, but will it cause 
> >> trouble if the prctl() was indiscriminately used by some applications, 
> >> will we get fake data?
> > 
> > It's their problem: overusing it will increase their CPU overhead. The two 
> > boundary worst-cases are that they either call it too frequently or too 
> > rarely:
> > 
> >  - too frequently: it approximates the current cpu-runtime work metric
> > 
> >  - too infrequently: we just ignore it and fall back to a runtime metric
> >    if it does not change.
> > 
> > It's not like it can be used to get preferential treatment - we don't ever 
> > balance other tasks against these tasks based on work throughput, we try 
> > to maximize this workload's work throughput.
> > 
> > What could happen is if an app is 'optimized' for a buggy scheduler by 
> > changing the work metric frequency. We offer no guarantee - apps will be 
> > best off (and users will be least annoyed) if apps honestly report their 
> > work metric.
> > 
> > Instrumentation/stats/profiling will also double check the correctness of 
> > this data: if developers/users start relying on the work metric as a 
> > substitute benchmark number, then app writers will have an additional 
> > incentive to make them correct.
> 
> I see, I could not figure out how to wisely using the info currently, 
> but I have the feeling that it will make scheduler very different ;-)
> 
> May be we could implement the API and get those info ready firstly 
> (along with the new sched-pipe which provide work tick info), then think 
> about the way to use them in scheduler, is there any patches on the way?

Absolutely.

Beyond the new prctl no new API is needed: a perf soft event could be 
added, and/or a tracepoint. Then perf stat and perf record could be used 
with it. 'perf bench' could be extended to generate the work tick in its 
'perf bench sched ...' workloads - and for 'perf bench mem numa' as well.

vsyscall-accelerating it could be a separate, more complex step: it needs 
a per thread writable vsyscall data area to make the overhead to 
applications near zero. Performance critical apps won't call an extra 
syscall.

Thanks,

	Ingo