Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Sun, 30 Aug 2015 17:31:39 +0200
From: Jens Gustedt <jens.gustedt@...ia.fr>
To: musl@...ts.openwall.com
Subject: Re: spin strategy in __wait

Am Samstag, den 29.08.2015, 20:57 -0400 schrieb Rich Felker:
> On Sat, Aug 29, 2015 at 09:14:17PM +0200, Jens Gustedt wrote:
> > The question that I asked myself is how much spinning to we need in
> > such lock/wait primitives. In the first attached graphic, the three
> > bottom curves show what difference three spinning strategies can make
> > for this test on a 2x2 hyperthreaded core machine. The bottom is no
> > spinning at all, the next is the current strategy implemented in musl,
> > and the third is a modification of that. As you can see, for a high
> > load of threads they can make a substantial difference, but you also
> > see that musl's actually strategy is not very different from doing no
> > spinning at all.
> 
> That's intentional when you have waiters. Notice in your graph that
> for the range 2-4 threads, performance in musl is comparable to your
> "no-shortcut" approach. This shows that the spinning is working as
> intended.

In that range, yes. When the machine is overcommitted with, say, 2 times
more threads than the number of cores, it is significantly off.

> BTW I think the names you chose are misleading. :-) Refraining to spin
> is not a "shortcut" musl is taking. It's not trying to be quicker, but
> to refrain from aggressive behavior which should lead to pathological
> lock unfairness. I would call the approaches aggressive-spin,
> opportunistic-spin, and no-spin.

Probably, they were really just "working" names for my scripts.

> > The "new" strategy is simply to avoid to take the shortcut that the
> > actual code takes when spinning. Here, currently when we detect that
> > are already waiters, we stop spinning and try to go into futex_wait
> > immediately. Where this sounds nice at a first glance, the figures
> > seem to indicate that this is not a good idea and that we would be
> > better off without. I'll send a patch for that in a next mail, and
> > also one that lets you modify the number of spins easily.
> 
> I don't think you can say we'd be better off in general with a single
> test.

Sure, but it should be taken as a data point. Average behavior of the
malloc implementation seems to be important to me.

> In areas like this you need to be able to demonstrate that there
> are no cases where the expected pathologically bad behavior can
> happen. I suspect you get lucky with malloc because there's a fair
> amount of code being executed between successive attempts on the lock
> rather than intense hammering.

Perhaps. But the behavior that I observe is not so much different
where I used these things as locks for the atomic stuff. There the
work that is done inside the critical section is just a memcpy/memcmp
of 16 byte.

> On the other hand, are you even counting the number of times each
> individual thread performs a malloc/free cycle, or only the total
> across all threads? Is it possible that, among your 256 threads, the
> vast majority of malloc/free cycles are being performed by a small
> number and the rest are stuck repeating futex_wait over and over?

I now also collected these statistics, and, no, all threads get their
share. I attach to data files (tab separated values) that show
this. These files have one line per run (10 sec each), with the
following columns

number of threads
average number per thread of calls to malloc/free during the 10 secs
standard deviation of that quantity
the min value over all threads
the max value over all threads

As you may see, the variation is the biggest around 3-4 threads, but
stays reasonable, I think, from overall min to overall max there is a
difference of 50% or so. Nobody is blocked.

This variation then gets really small for up to 64 threads. For 128
and 256 it is a bit bigger, but overall the numbers are a bit too
small for that cases and one should perhaps run the test for more than
10 secs.

> > Now the only situation I thought of where this could be important is
> > monoprocessors where actually spinning might not be so good and
> > aborting it early good be necessary. So I rand the same test with
> > taskset to nail the process to just one core. The result of that are
> > the top two curves. As you can see, here the spinning strategy has
> > almost no influence so I think we are safe to apply this patch.
> 
> This is very surprising to me, which further makes me doubt the test
> methodology. The current number of spins is tuned to be comparable in
> magnitude to the syscall cost, at least on x86.

On my machine, x86_64, with are quite recent kernel it is not, it is a
factor of 10 off. As I said, I observe a factor of 10 between a spin
and a syscall, and not 100 as the existing code suggests.

And, BTW, I also observed that the call to "a_spin" doesn't have much
effect on my machine, neither positive nor negative.

> Adding this extra cost
> to each contended lock operation (you basically never stop spinning
> early on single-core) should produce a measurable difference.

No, on single core you have only one thread active at any time, and
the probability that it is descheduled in the middle when holding the
lock is tiny. So basically nobody is ever spinning as long as the time
that is spend in the critical section is small. (And you may have
sensible outlayers, once a thread is descheduled in the middle.)

> > Now all of this can also be read as a performance test of the malloc
> > subsystem, and here my feeling goes in the direction that Rich
> > recently indicated. The performance of the "application" is much
> > better if I eliminate all parallelism. As an additional indication
> > there are to additional curves that fix the process to one core and
> > its hyperthreaded sibling.
> > 
> > So maybe we might be better off with a simpler malloc strategy that
> > serializes all requests.
> 
> I tried this a few weeks ago and it performed significantly worse on
> my malloc stress tests. And it will get even worse as we try to scale
> to larger numbers of cores, or worse yet NUMA...

Did you try running your malloc stress test with taskset to nail the
process to one core?

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 ::




View attachment "test-16b-spin0-per-thread.csv" of type "text/csv" (2623 bytes)

View attachment "test-16b-no-shortcut-per-thread.csv" of type "text/csv" (2595 bytes)

Download attachment "signature.asc" of type "application/pgp-signature" (182 bytes)

Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.