Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1757460AbYLFJzz (ORCPT ); Sat, 6 Dec 2008 04:55:55 -0500 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1752812AbYLFJzr (ORCPT ); Sat, 6 Dec 2008 04:55:47 -0500 Received: from fk-out-0910.google.com ([209.85.128.191]:11504 "EHLO fk-out-0910.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752753AbYLFJzq (ORCPT ); Sat, 6 Dec 2008 04:55:46 -0500 DomainKey-Signature: a=rsa-sha1; c=nofws; d=gmail.com; s=gamma; h=message-id:date:from:user-agent:mime-version:to:cc:subject :references:in-reply-to:content-type:content-transfer-encoding; b=cFSqfADc5vwZa3sB5p4t7aVLV7ZaJxjbqtdpKGCVXdqcqmT+vVIRsX+6QKLMKaYpft 1I7zmQVcxQwTfAUYww1+s1pE7tNSUdpjnsS+gTpkj6RQ89IpoMP58vyTsMPtdsUH9B3K zi+wPM1aLxKj+eNBgF4ahkBPbeO4sE/CgRvfs= Message-ID: <493A4C1C.3020102@gmail.com> Date: Sat, 06 Dec 2008 11:55:40 +0200 From: =?ISO-8859-1?Q?T=F6r=F6k_Edwin?= User-Agent: Mozilla-Thunderbird 2.0.0.17 (X11/20081018) MIME-Version: 1.0 To: Ying Han CC: linux-mm@kvack.org, linux-kernel , akpm , Ingo Molnar , Mike Waychison , David Rientjes , Rohit Seth , Hugh Dickins , Peter Zijlstra , "H. Peter Anvin" , Lee Schermerhorn , Nick Piggin Subject: Re: [RFC v2][PATCH]page_fault retry with NOPAGE_RETRY References: <604427e00812051140s67b2a89dm35806c3ee3b6ed7a@mail.gmail.com> <493A4B48.1050706@gmail.com> In-Reply-To: <493A4B48.1050706@gmail.com> Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: 8bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 4256 Lines: 89 On 2008-12-06 11:52, T?r?k Edwin wrote: > On 2008-12-05 21:40, Ying Han wrote: > >> changelog[v2]: >> - reduce the runtime overhead by extending the 'write' flag of >> handle_mm_fault() to indicate the retry hint. >> - add another two branches in filemap_fault with retry logic. >> - replace find_lock_page with find_lock_page_retry to make the code >> cleaner. >> >> todo: >> - there is potential a starvation hole with the retry. By the time the >> retry returns, the pages might be released. we can make change by holding >> page reference as well as remembering what the page "was"(in case the >> file was truncated). any suggestion here are welcomed. >> >> I also made patches for all other arch. I am posting x86_64 here first and >> i will post others by the time everyone feels comfortable of this patch. >> >> Edwin, please test this patch with your testcase and check if you get any >> performance improvement of mmap over read. I added another two more places >> in filemap_fault with retry logic which you might hit in your privous >> experiment. >> >> > > I get much better results with this patch than with v1, thanks! > > mmap now scales almost as well as read does (there is a small ~5% > overhead), which is a significant improvement over not scaling at all! > > Here are the results when running my testcase: > > Number of threads ->, 1,,, 2,,, 4,,, 8,,, 16 > Kernel version, read, mmap, mixed, read, mmap, mixed, read, mmap, mixed, > read, mmap, mixed, read, mmap, mixed > 2.6.28-rc7-tip, 27.55, 26.18, 27.06, 16.18, 16.97, 16.10, 11.06, 11.64, > 11.41, 9.38, 9.97, 9.31, 9.37, 9.82, 9.3 > > > Here are the /proc/lock_stat output when running my testcase, contention > is lower (34911+10462 vs 58590+7231), and waittime-total is better > (57 601 464 vs 234 170 024) > > lock_stat version 0.3 > ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- > class name con-bounces contentions > waittime-min waittime-max waittime-total acq-bounces > acquisitions holdtime-min holdtime-max holdtime-total > ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- > &mm->mmap_sem-W: 5843 > 10462 2.89 138824.72 14217159.52 > 18965 84205 1.81 5031.07 725293.65 > &mm->mmap_sem-R: 20208 > 34911 4.87 136797.26 57601464.49 55797 > 1110394 1.89 164918.52 30551371.71 > --------------- > &mm->mmap_sem 5341 > [] sys_munmap+0x47/0x80 > &mm->mmap_sem 28579 > [] do_page_fault+0x172/0xab0 > &mm->mmap_sem 5030 > [] sys_mmap+0xf1/0x140 > &mm->mmap_sem 6331 > [] find_lock_page_retry+0xde/0xf0 > --------------- > &mm->mmap_sem 13558 > [] find_lock_page_retry+0xde/0xf0 > &mm->mmap_sem 4694 > [] sys_munmap+0x47/0x80 > &mm->mmap_sem 3681 > [] sys_mmap+0xf1/0x140 > &mm->mmap_sem 23374 > [] do_page_fault+0x172/0xab0 > > > On clamd: > > Here holdtime-total is better (1 493 154 + 2 395 987 vs 2 087 538 + 2 > 514 673), and number of contentions on read > (458 052 vs 5851 typo, should have been: 458 052 vs 585 119 -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/