Re: [RFC PATCH 0/2] ipc: do not hold ipc lock more than necessary

[Davidlohr Bueso]

On Fri, 2013-03-01 at 17:32 -0800, Linus Torvalds wrote:
> On Fri, Mar 1, 2013 at 4:16 PM, Davidlohr Bueso <davidlohr.bueso@xxxxxx> wrote:
> >
> > With Rik's semop-multi.c microbenchmark we can see the following
> > results:
> 
> Ok, that certainly looks very good.
> 
> > +  59.40%            a.out  [kernel.kallsyms]  [k] _raw_spin_lock
> > +  17.47%            a.out  [kernel.kallsyms]  [k] _raw_spin_lock
> 
> I had somewhat high expectations, but that's just better than I really
> hoped for. Not only is the percentage down, it's down for the case of
> a much smaller number of overall cycle cost, so it's a really big
> reduction in contention spinning.
> 
> Of course, contention will come back and overwhelm you at *some*
> point, but it seems the patches certainly moved the really bad
> contention point out some way..
> 
> > +   6.14%            a.out  [kernel.kallsyms]  [k] sys_semtimedop
> > +  11.08%            a.out  [kernel.kallsyms]  [k] sys_semtimedop
> > While the _raw_spin_lock time is drastically reduced, others do increase.
> > This results in an overall speedup of ~1.7x regarding ops/sec.
> 
> Actually, the others don't really increase. Sure, the *percentages* go
> up, but that's just because it has to add up to 100% in the end. So
> it's not that you're moving costs from one place to another - the 1.7x
> speedup is the real reduction in costs, and then that 6.14% -> 11.08%
> "growth" is really nothing but that (and yes, 1.7 x 6.14 really does
> get pretty close).
> 
> So nothing really got slower, despite the percentages going up.
> 
> Looks good to me. Of course, the *real* issue is if this is a win on
> real code too. And I bet it is, it just won't be quite as noticeable.
> But if anything, real code is likely to have less contention to begin
> with, because it has more things going on outside of the spinlocks. So
> it should see an improvement, but not nearly the kind of improvement
> you quote here.
> 
> Although your 800-user swingbench numbers were pretty horrible, so
> maybe that case can improve by comparable amounts in the bad cases.
> 

Absolutely, I'll be sure to try these changes with my Oracle workloads
and report with some numbers. This obviously still needs a lot of
testing.

Thanks,
Davidlohr

--
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/