From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (qmail 45453 invoked by alias); 4 Oct 2016 09:45:17 -0000 Mailing-List: contact gcc-patches-help@gcc.gnu.org; run by ezmlm Precedence: bulk List-Id: List-Archive: List-Post: List-Help: Sender: gcc-patches-owner@gcc.gnu.org Received: (qmail 45429 invoked by uid 89); 4 Oct 2016 09:45:16 -0000 Authentication-Results: sourceware.org; auth=none X-Virus-Found: No X-Spam-SWARE-Status: No, score=3.5 required=5.0 tests=AWL,BAYES_00,FREEMAIL_FROM,RCVD_IN_DNSWL_NONE,SPAM_BODY,SPF_PASS autolearn=no version=3.3.2 spammy=11.6, 3.6, 15.5, HX-Envelope-From:sk:richard X-HELO: mail-wm0-f66.google.com Received: from mail-wm0-f66.google.com (HELO mail-wm0-f66.google.com) (74.125.82.66) by sourceware.org (qpsmtpd/0.93/v0.84-503-g423c35a) with ESMTP; Tue, 04 Oct 2016 09:45:04 +0000 Received: by mail-wm0-f66.google.com with SMTP id f193so14136637wmg.2 for ; Tue, 04 Oct 2016 02:45:04 -0700 (PDT) X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20130820; h=x-gm-message-state:mime-version:in-reply-to:references:from:date :message-id:subject:to:cc:content-transfer-encoding; bh=FIxaWKhFglzLdf8c0uOyOCSkvgX73t+V/aPXzCV4YK8=; b=g6XSEwPeEu7FE60tldba4WArSZ7nM6lCTmGllFSqXtLw5Sb7n8IyKqZhzZ237qklMW bYv/lJ303bttwj5iNkFZAmqAxSAez+ACwLVBGrZdl+mLMXFtQ84ulza1zGEuKkHcwVm4 odwOkD97h21f6No9LpOSWDZ+UCBdyTSMshfp/sSWupQcqWQxOuMZXKMdBEL9tQht9Kul sBKNmMo1SM2RoDu8Yvu+5kd1BIZJXX5u9xSl3c8arv5SsHt13MTDMTq2QfHxkG6QLtZk zsYZ9fHIgM73lYYUbrxtxxH0pBpbFIguxm/IMEYDhaUaKTBb0UyamUMYJxl1Swufq5YV 4FYQ== X-Gm-Message-State: AA6/9RlyCYNfKHap7zozCnOYdS+R1FqbihjIRLKWEMnblU+56FF/X44WkhQuqyM2Xs0VoTlvgrXTYJKPVqYNvg== X-Received: by 10.194.59.84 with SMTP id x20mr2168467wjq.89.1475574302712; Tue, 04 Oct 2016 02:45:02 -0700 (PDT) MIME-Version: 1.0 Received: by 10.28.155.210 with HTTP; Tue, 4 Oct 2016 02:45:02 -0700 (PDT) In-Reply-To: <8932e842-2457-64a0-76bc-a81b9a9a9b31@suse.cz> References: <1253ac69-3301-f185-e43a-a34cadf8f51e@suse.cz> <67fda6d2-9b3e-a0d1-effc-34e1115030b2@acm.org> <1ff3cc75-7cee-79f3-395b-ef7a4d286a3d@acm.org> <04a05835-4666-4d7d-c1a9-d4bcc4ea924a@suse.cz> <87k2fpdatl.fsf@tassilo.jf.intel.com> <6f8b1905-818b-bfff-1bf3-5ba04f3b4b64@suse.cz> <20160818155130.GE5871@two.firstfloor.org> <20160818155449.GP14857@tucnak.redhat.com> <5798a459-2fc7-d82a-f89b-30a45a03c831@suse.cz> <8932e842-2457-64a0-76bc-a81b9a9a9b31@suse.cz> From: Richard Biener Date: Tue, 04 Oct 2016 09:45:00 -0000 Message-ID: Subject: Re: [RFC] Speed-up -fprofile-update=atomic To: =?UTF-8?Q?Martin_Li=C5=A1ka?= Cc: Jakub Jelinek , Andi Kleen , Jeff Law , Nathan Sidwell , GCC Patches , "Hubicha, Jan" Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: quoted-printable X-IsSubscribed: yes X-SW-Source: 2016-10/txt/msg00139.txt.bz2 On Thu, Sep 15, 2016 at 12:00 PM, Martin Li=C5=A1ka wrote: > On 09/07/2016 02:09 PM, Richard Biener wrote: >> On Wed, Sep 7, 2016 at 1:37 PM, Martin Li=C5=A1ka 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 iss= ues (short lived threads, so we need to somehow re-use counters from termin= ated threads, large number of threads and thus using too much memory for th= e counters) >>>> >>>> Richard. >>> >>> Hello. >>> >>> I've got written the approach on my TODO list, let's see whether it wou= ld be doable in a reasonable amount of time. >>> >>> I've just finished some measurements to illustrate slow-down of -fprofi= le-update=3Datomic approach. >>> All numbers are: no profile, -fprofile-generate, -fprofile-generate -fp= rofile-update=3Datomic >>> 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 no= t 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 upda= te >> 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 nu= mber >> 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=3Datomic, or to add a new= option value > like -fprofile-update=3Datomic-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? Hmm, I thought I suggested that you can simply use automatic storage (which effectively is TLS...) for regions that are not forked or abnormally left (which means SESE regions that have no calls that eventually terminate or throw externally). So why did you end up with TLS? Richard. > Martin > >> >>> Martin >>> >>>> >>>>> Jakub >>>> >>>> >>> >