Re: [PATCH V2 0/6][RFC] futex: FUTEX_LOCK with optional adaptivespinning

[Darren Hart]

Darren Hart wrote:
> Avi Kivity wrote:
>
> > > > At 10%
> > > > duty cycle you have 25 waiters behind the lock on average.  I don't 
> > > > think this is realistic, and it means that spinning is invoked only 
> > > > rarely.
> > >
> > > Perhaps some instrumentation is in order, it seems to get invoked 
> > > enough to achieve some 20% increase in lock/unlock iterations. 
> > > Perhaps another metric would be of more value - such as average wait 
> > > time?
> >
> > Why measure an unrealistic workload?
>
> No argument there, thus my proposal for an alternate configuration below.
>
> > > > I'd be interested in seeing runs where the average number of waiters 
> > > > is 0.2, 0.5, 1, and 2, corresponding to moderate-to-bad contention.
> > > > 25 average waiters on compute bound code means the application needs 
> > > > to be rewritten, no amount of mutex tweaking will help it.
> > >
> > > Perhaps something NR_CPUS threads would be of more interest? 
> >
> > That seems artificial.
>
> How so? Several real world applications use one thread per CPU to 
> dispatch work to, wait for events, etc.
>
> > > At 10% that's about .8 and at 25% the 2 of your upper limit. I could 
> > > add a few more duty-cycle points and make 25% the max. I'll kick that 
> > > off and post the results... probably tomorrow, 10M iterations takes a 
> > > while, but makes the results relatively stable.
> >
> > Thanks.  But why not vary the number of threads as well?
>
> Absolutely, I don't disagree that all the variables should vary in order 
> to get a complete picture. I'm starting with 8 - it takes several hours 
> to collect the data.

While this might be of less interest after today's discussion, I 
promised to share the results of a run with 8 threads with a wider 
selection of lower duty-cycles. The results are very poor for adaptive 
and worse for aas (multiple spinners) compared to normal FUTEX_LOCK. As 
Thomas and Peter have pointed out, the implementation is sub-optimal. 
Before abandoning this approach I will see if I can find the bottlenecks 
and simplify the kernel side of things. My impression is that I am doing 
a lot more work in the kernel, especially in the adaptive loop, than is 
really necessary.

Both the 8 and 256 Thread plots can be viewed here:

http://www.kernel.org/pub/linux/kernel/people/dvhart/adaptive_futex/v4/

--
Darren Hart
IBM Linux Technology Center
Real-Time Linux Team
--
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/