From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-3.5 required=3.0 tests=BAYES_00,DKIM_INVALID, DKIM_SIGNED,HEADER_FROM_DIFFERENT_DOMAINS,MAILING_LIST_MULTI,SPF_HELO_NONE, SPF_PASS autolearn=no autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id C438EC433E1 for ; Wed, 19 Aug 2020 15:38:08 +0000 (UTC) Received: from lists.gnu.org (lists.gnu.org [209.51.188.17]) (using TLSv1.2 with cipher ECDHE-RSA-AES256-GCM-SHA384 (256/256 bits)) (No client certificate requested) by mail.kernel.org (Postfix) with ESMTPS id 8AF22207FF for ; Wed, 19 Aug 2020 15:38:08 +0000 (UTC) Authentication-Results: mail.kernel.org; dkim=fail reason="signature verification failed" (2048-bit key) header.d=igalia.com header.i=@igalia.com header.b="A64VMJg4" DMARC-Filter: OpenDMARC Filter v1.3.2 mail.kernel.org 8AF22207FF Authentication-Results: mail.kernel.org; dmarc=none (p=none dis=none) header.from=igalia.com Authentication-Results: mail.kernel.org; spf=pass smtp.mailfrom=qemu-devel-bounces+qemu-devel=archiver.kernel.org@nongnu.org Received: from localhost ([::1]:59654 helo=lists1p.gnu.org) by lists.gnu.org with esmtp (Exim 4.90_1) (envelope-from ) id 1k8QAB-0004ZV-OC for qemu-devel@archiver.kernel.org; Wed, 19 Aug 2020 11:38:07 -0400 Received: from eggs.gnu.org ([2001:470:142:3::10]:37872) by lists.gnu.org with esmtps (TLS1.2:ECDHE_RSA_AES_256_GCM_SHA384:256) (Exim 4.90_1) (envelope-from ) id 1k8Q9P-00048x-Pb; Wed, 19 Aug 2020 11:37:20 -0400 Received: from fanzine.igalia.com ([178.60.130.6]:55339) by eggs.gnu.org with esmtps (TLS1.2:ECDHE_RSA_AES_128_GCM_SHA256:128) (Exim 4.90_1) (envelope-from ) id 1k8Q9M-00027Y-Cm; Wed, 19 Aug 2020 11:37:19 -0400 DKIM-Signature: v=1; a=rsa-sha256; q=dns/txt; c=relaxed/relaxed; d=igalia.com; s=20170329; h=Content-Type:MIME-Version:Message-ID:Date:References:In-Reply-To:Subject:Cc:To:From; bh=0T4RlsT2sSUKUGPqYvw5XiBU2ZIeTCZIyh7Mpg1z6xI=; b=A64VMJg41OeI27A+16EhWaHLt1RfYF9P8k4elxmZQ8UjXMo36wCjTfbtR2t6ijbvDL863ipBZI9fS0iXADkqORFHnx0KlciMFs2rNK3LzGuoouNVnkkyD0qAbLuyVXIVPeYCZFXnLt9QuiyTQFXSqKJNTM6lkEKkp14NL8wKK7lfdUICbpKjfdEWpYHxY5yofrRYf01SQbK0NnuKP05Mq6q4WT2XKS6yiaIExhQWmY+rxsqyvQWqhouxoyjk/wfPOCM4gxlB6ivpd+kcIxa7tLVT5K86K3rjzzVIpWE9ttrPmtgQXxeZax6be7DS2HsULm5CJZ3GGDi29pslakfeHA==; Received: from maestria.local.igalia.com ([192.168.10.14] helo=mail.igalia.com) by fanzine.igalia.com with esmtps (Cipher TLS1.2:ECDHE_RSA_AES_128_GCM_SHA256:128) (Exim) id 1k8Q9I-0002oF-OA; Wed, 19 Aug 2020 17:37:12 +0200 Received: from berto by mail.igalia.com with local (Exim) id 1k8Q9I-0002Na-Et; Wed, 19 Aug 2020 17:37:12 +0200 From: Alberto Garcia To: Kevin Wolf Subject: Re: [PATCH 0/1] qcow2: Skip copy-on-write when allocating a zero cluster In-Reply-To: <20200819150711.GE10272@linux.fritz.box> References: <20200817101019.GD11402@linux.fritz.box> <20200817155307.GS11402@linux.fritz.box> <20200819150711.GE10272@linux.fritz.box> User-Agent: Notmuch/0.18.2 (http://notmuchmail.org) Emacs/24.4.1 (i586-pc-linux-gnu) Date: Wed, 19 Aug 2020 17:37:12 +0200 Message-ID: MIME-Version: 1.0 Content-Type: text/plain Received-SPF: pass client-ip=178.60.130.6; envelope-from=berto@igalia.com; helo=fanzine.igalia.com X-detected-operating-system: by eggs.gnu.org: First seen = 2020/08/19 10:25:17 X-ACL-Warn: Detected OS = Linux 2.2.x-3.x (no timestamps) [generic] [fuzzy] X-Spam_score_int: -20 X-Spam_score: -2.1 X-Spam_bar: -- X-Spam_report: (-2.1 / 5.0 requ) BAYES_00=-1.9, DKIM_SIGNED=0.1, DKIM_VALID=-0.1, DKIM_VALID_AU=-0.1, DKIM_VALID_EF=-0.1, SPF_HELO_NONE=0.001, SPF_PASS=-0.001, URIBL_BLOCKED=0.001 autolearn=ham autolearn_force=no X-Spam_action: no action X-BeenThere: qemu-devel@nongnu.org X-Mailman-Version: 2.1.23 Precedence: list List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Cc: bfoster@redhat.com, Vladimir Sementsov-Ogievskiy , qemu-devel@nongnu.org, qemu-block@nongnu.org, Max Reitz Errors-To: qemu-devel-bounces+qemu-devel=archiver.kernel.org@nongnu.org Sender: "Qemu-devel" On Wed 19 Aug 2020 05:07:11 PM CEST, Kevin Wolf wrote: >> I checked with xfs on my computer. I'm not very familiar with that >> filesystem so I was using the default options and I didn't tune >> anything. >> >> What I got with my tests (using fio): >> >> - Using extent_size_hint didn't make any difference in my test case (I >> do see a clear difference however with the test case described in >> commit ffa244c84a). > > Hm, interesting. What is your exact fio configuration? Specifically, > which iodepth are you using? I guess with a low iodepth (and O_DIRECT), > the effect of draining the queue might not be as visible. fio --filename=/dev/vdb --direct=1 --randrepeat=1 --eta=always --ioengine=libaio --iodepth=32 --numjobs=1 --name=test --size=25G --io_limit=25G --ramp_time=5 --rw=randwrite --bs=4k --runtime=60 >> - preallocation=off is still faster than preallocation=metadata. > > Brian, can you help us here with some input? > > Essentially what we're having here is a sparse image file on XFS that > is opened with O_DIRECT (presumably - Berto, is this right?), and > Berto is seeing cases where a random write benchmark is faster if > we're doing the 64k ZERO_RANGE + 4k pwrite when touching a 64k cluster > for the first time compared to always just doing the 4k pwrite. This > is with a 1 MB extent size hint. A couple of notes: - Yes, it's O_DIRECT (the image is opened with cache=none and fio uses --direct=1). - The extent size hint is the default one, I didn't change or set anything for this test (or should I have?). > From the discussions we had the other day [1][2] I took away that your > suggestion is that we should not try to optimise things with > fallocate(), but just write the areas we really want to write and let > the filesystem deal with the sparse parts. Especially with the extent > size hint that we're now setting, I'm surprised to hear that doing a > ZERO_RANGE first still seems to improve the performance. > > Do you have any idea why this is happening and what we should be doing > with this? > > [1] https://bugzilla.redhat.com/show_bug.cgi?id=1850660 > [2] https://bugzilla.redhat.com/show_bug.cgi?id=1666864 > >> If I disable handle_alloc_space() (so there is no ZERO_RANGE used) >> then it is much slower. > > This makes some sense because then we're falling back to writing > explicit zero buffers (unless you disabled that, too). Exactly, this happens on both ext4 and xfs. >> - With preallocation=falloc I get the same results as with >> preallocation=metadata. > > Interesting, this means that the fallocate() call costs you basically > no time. I would have expected preallocation=falloc to be a little > faster. I would expect preallocation=falloc to be at least as fast as preallocation=off (and it is, on ext4). However on xfs it seems to be slower (?). It doesn't make sense to me. >> - preallocation=full is the fastest by far. > > I guess this saves the conversion of unwritten extents to fully > allocated ones? However it is *much* *much* faster. I assume I must be missing something on how the filesystem works. I ran the test again on a newly created filesystem just to make sure, here are the full results (numbers are IOPS): |----------------------+-------+-------| | preallocation | ext4 | xfs | |----------------------+-------+-------| | off | 11688 | 6981 | | off (w/o ZERO_RANGE) | 2780 | 3196 | | metadata | 9132 | 5764 | | falloc | 13108 | 5727 | | full | 16351 | 40759 | |----------------------+-------+-------| Berto