On 09/07/2016 02:09 PM, Richard Biener wrote: > On Wed, Sep 7, 2016 at 1:37 PM, Martin Liška wrote: >> On 08/18/2016 06:06 PM, Richard Biener wrote: >>> On August 18, 2016 5:54:49 PM GMT+02:00, Jakub Jelinek wrote: >>>> On Thu, Aug 18, 2016 at 08:51:31AM -0700, Andi Kleen wrote: >>>>>> I'd prefer to make updates atomic in multi-threaded applications. >>>>>> The best proxy we have for that is -pthread. >>>>>> >>>>>> Is it slower, most definitely, but odds are we're giving folks >>>>>> garbage data otherwise, which in many ways is even worse. >>>>> >>>>> It will likely be catastrophically slower in some cases. >>>>> >>>>> Catastrophically as in too slow to be usable. >>>>> >>>>> An atomic instruction is a lot more expensive than a single >>>> increment. Also >>>>> they sometimes are really slow depending on the state of the machine. >>>> >>>> Can't we just have thread-local copies of all the counters (perhaps >>>> using >>>> __thread pointer as base) and just atomically merge at thread >>>> termination? >>> >>> I suggested that as well but of course it'll have its own class of issues (short lived threads, so we need to somehow re-use counters from terminated threads, large number of threads and thus using too much memory for the counters) >>> >>> Richard. >> >> Hello. >> >> I've got written the approach on my TODO list, let's see whether it would be doable in a reasonable amount of time. >> >> I've just finished some measurements to illustrate slow-down of -fprofile-update=atomic approach. >> All numbers are: no profile, -fprofile-generate, -fprofile-generate -fprofile-update=atomic >> c-ray benchmark (utilizing 8 threads, -O3): 1.7, 15.5., 38.1s >> unrar (utilizing 8 threads, -O3): 3.6, 11.6, 38s >> tramp3d (1 thread, -O3): 18.0, 46.6, 168s >> >> So the slow-down is roughly 300% compared to -fprofile-generate. I'm not having much experience with default option >> selection, but these numbers can probably help. >> >> Thoughts? > > Look at the generated code for an instrumented simple loop and see that for > the non-atomic updates we happily apply store-motion to the counter update > and thus we only get one counter update per loop exit rather than one per > loop iteration. Now see what happens for the atomic case (I suspect you > get one per iteration). > > I'll bet this accounts for most of the slowdown. > > Back in time ICC which had atomic counter updates (but using function > calls - ugh!) had a > 1000% overhead with FDO for tramp3d (they also > didn't have early inlining -- removing abstraction helps reducing the number > of counters significantly). > > Richard. Hi. During Cauldron I discussed with Richi approaches how to speed-up ARCS profile counter updates. My first attempt is to utilize TLS storage, where every function is accumulating arcs counters. These are eventually added (using atomic operations) to the global one at the very end of a function. Currently I rely on target support of TLS, which is questionable whether to have such a requirement for -fprofile-update=atomic, or to add a new option value like -fprofile-update=atomic-tls? Running the patch on tramp3d, compared to previous numbers, it takes 88s to finish. Time shrinks to 50%, compared to the current implementation. Thoughts? Martin > >> Martin >> >>> >>>> Jakub >>> >>> >>