Rich,
I hope with the graphics at hand, you'll understand a bit better what
I was up to. Again my apologies.

Still I try to answer to some of the arguments that you give below:

Am Samstag, den 29.08.2015, 13:16 -0400 schrieb Rich Felker:
> On Sat, Aug 29, 2015 at 10:50:44AM +0200, Jens Gustedt wrote:
> > Remove a test in __wait that looked if other threads already attempted to
> > go to sleep in futex_wait.
> > 
> > This has no impact on the fast path. But other than one might think at a
> > first glance, this slows down if there is congestion.
> > 
> > Applying this patch shows no difference in behavior in a mono-core
> > setting, so it seems that this shortcut is just superfluous.
> 
> The purpose of this code is is twofold: improving fairness of the lock
> and avoiding burning cpu time that's _known_ to be a waste.
> 
> If you spin on a lock that already has waiters, the thread that spins
> has a much better chance to get the lock than any of the existing
> waiters which are in futex_wait. Assuming sufficiently many cores that
> all threads that are not sleeping don't get preempted, the spinning
> thread is basically guaranteed to get the lock unless it spins so long
> to make it futex_wait. This is simply because returning from
> futex_wake (which all the other waiters have to do) takes a lot more
> time than one spin. I suspect there are common loads under which many
> of the waiters will NEVER get the lock.

Yes and no. I benched the things to know a bit more. On my machine one
loop in a spin lock is just about 10 times faster than a failed call
to futex_wait.

Even for the current strategy, one of the futex_waiting threads gets
woken up and gets his chance with a new spinning phase.

So the difference isn't dramatic, just one order of magnitude and
everybody gets his chance. These chances are not equal, sure, but
NEVER in capitals is certainly a big word.

On the other hands the difference in throughput on the multi-core
setting between the different spin versions is dramatic for malloc, I
find.

> The other issue is simply wasted cpu time. Unlike in the no-waiters
> case where spinning _can_ save a lot of cpu time (no futex_wait or
> futex_wake syscall), spinning in the with-waiters case always wastes
> cpu time. If the spinning thread gets the lock when the old owner
> unlocks it, the old owner has already determined that it has to send a
> futex_wake. So now you have (1) the wasted futex_wake syscall in the
> old owner thread and (2) a "spurious" wake in one of the waiters,
> which wakes up only to find that the lock was already stolen, and
> therefore has to call futex_wait again and start the whole process
> over.

This wasted CPU time was one of my worries at first, this is why I
made the mono core test. Basically on mono core spinning is *always* a
waste, and going into futex_wait immediately should always be a
win. (BTW, the current implementation should have a performance gap
for exactly 2 threads that alternate on the lock.)

But, there is almost no difference between the two strategies, the
bias even here goes a tiny bit towards doing some spin, even under
very high load. So, no, I don't think there is much wasted CPU
time. At least performance seems to be largely dominated by other
things, so "spinning" shouldn't be a worry for that in the
current state of things.


Jens

-- 
:: INRIA Nancy Grand Est ::: Camus ::::::: ICube/ICPS :::
:: ::::::::::::::: office Strasbourg : +33 368854536   ::
:: :::::::::::::::::::::: gsm France : +33 651400183   ::
:: ::::::::::::::: gsm international : +49 15737185122 ::
:: http://icube-icps.unistra.fr/index.php/Jens_Gustedt ::