From: Jiaying Zhang Subject: Re: ext4 DIO read performance issue on SSD Date: Wed, 14 Oct 2009 14:42:51 -0700 Message-ID: <5df78e1d0910141442g680edac9m6bce0f9eb21f8ea6@mail.gmail.com> References: <5df78e1d0910091634q22e6a372g3738b0d9e9d0e6c9@mail.gmail.com> <1255546117.4377.62.camel@mingming-laptop> <5df78e1d0910141248h1f537863n97991585e6147ca7@mail.gmail.com> <1255553852.4377.63.camel@mingming-laptop> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1 Content-Transfer-Encoding: QUOTED-PRINTABLE Cc: ext4 development , Michael Rubin , Manuel Benitez , Andrew Morton To: Mingming Return-path: Received: from smtp-out.google.com ([216.239.33.17]:36248 "EHLO smtp-out.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752048AbZJNVoE convert rfc822-to-8bit (ORCPT ); Wed, 14 Oct 2009 17:44:04 -0400 Received: from zps76.corp.google.com (zps76.corp.google.com [172.25.146.76]) by smtp-out.google.com with ESMTP id n9ELgssu024426 for ; Wed, 14 Oct 2009 22:42:55 +0100 Received: from yxe37 (yxe37.prod.google.com [10.190.2.37]) by zps76.corp.google.com with ESMTP id n9ELgpDI007066 for ; Wed, 14 Oct 2009 14:42:51 -0700 Received: by yxe37 with SMTP id 37so238458yxe.31 for ; Wed, 14 Oct 2009 14:42:51 -0700 (PDT) In-Reply-To: <1255553852.4377.63.camel@mingming-laptop> Sender: linux-ext4-owner@vger.kernel.org List-ID: On Wed, Oct 14, 2009 at 1:57 PM, Mingming wrote: > On Wed, 2009-10-14 at 12:48 -0700, Jiaying Zhang wrote: >> On Wed, Oct 14, 2009 at 11:48 AM, Mingming wrote: >> > On Fri, 2009-10-09 at 16:34 -0700, Jiaying Zhang wrote: >> >> Hello, >> >> >> >> Recently, we are evaluating the ext4 performance on a high speed = SSD. >> >> One problem we found is that ext4 performance doesn't scale well = with >> >> multiple threads or multiple AIOs reading a single file with O_DI= RECT. >> >> E.g., with 4k block size, multiple-thread DIO AIO random read on = ext4 >> >> can lose up to 50% throughput compared to the results we get via = RAW IO. >> >> >> >> After some initial analysis, we think the ext4 performance proble= m is caused >> >> by the use of i_mutex lock during DIO read. I.e., during DIO read= , we grab >> >> the i_mutex lock in __blockdev_direct_IO because ext4 uses the de= fault >> >> DIO_LOCKING from the generic fs code. I did a quick test by calli= ng >> >> blockdev_direct_IO_no_locking() in ext4_direct_IO() and I saw ext= 4 DIO read >> >> got 99% performance as raw IO. >> >> >> > >> > This is very interesting...and impressive number. >> > >> > I tried to change ext4 to call blockdev_direct_IO_no_locking() dir= ectly, >> > but then realize that we can't do this all the time, as ext4 suppo= rt >> > ext3 non-extent based files, and uninitialized extent is not suppo= rt on >> > ext3 format file. >> > >> >> As we understand, the reason why we want to take i_mutex lock dur= ing DIO >> >> read is to prevent it from accessing stale data that may be expos= ed by a >> >> simultaneous write. We saw that Mingming Cao has implemented a pa= tch set >> >> with which when a get_block request comes from direct write, ext4= only >> >> allocates or splits an uninitialized extent. That uninitialized e= xtent >> >> will be marked as initialized at the end_io callback. >> > >> > Though I need to clarify that with all the patches in mainline, we= only >> > treat new allocated blocks form direct io write to holes, not to w= rites >> > to the end of file. I actually have proposed to treat the write to= the >> > end of file also as unintialized extents, but there is some concer= ns >> > that this getting tricky with updating inode size when it is async= IO >> > direct IO. So it didn't getting done yet. >> > >> >> =A0We are wondering >> >> whether we can extend this idea to buffer write as well. I.e., we= always >> >> allocate an uninitialized extent first during any write and conve= rt it >> >> as initialized at the time of end_io callback. This will eliminat= e the need >> >> to hold i_mutex lock during direct read because a DIO read should= never get >> >> a block marked initialized before the block has been written with= new data. >> >> >> > >> > Oh I don't think so. For buffered IO, the data is being copied to >> > buffer, direct IO read would first flush what's in page cache to d= isk, >> >> Hmm, do you mean the filemap_write_and_wait_range() in >> __blockdev_direct_IO? > > yes, that's the one to flush the page cache before direct read. > I don't think that function is called with DIO_NO_LOCKING. Also, if we no longer hold i_mutex lock during dio read, I think there is a time window that a buffer write can allocate an initialize block after dio read flushes page cache but before it calls get_block. Then that dio read can get that initialized block with stale data. Jiaying >> Or do we flush page cache after calling >> get_block in dio read? >> >> Jiaying >> >> > then read from disk. So if there is concurrent buffered write and = direct >> > read, removing the i_mutex locks from the direct IO path should st= ill >> > gurantee the right order, without having to treat buffered allocat= ion >> > with uninitialized extent/end_io. >> > >> > The i_mutex lock, from my understanding, is there to protect direc= t IO >> > write to hole and concurrent direct IO read, we should able to rem= ove >> > this lock for extent based ext4 file. >> > >> >> >> >> We haven't implemented anything yet because we want to ask here f= irst to >> >> see whether this proposal makes sense to you. >> >> >> > >> > It does make sense to me. >> > >> > Mingming >> >> Regards, >> >> >> >> Jiaying >> >> -- >> >> To unsubscribe from this list: send the line "unsubscribe linux-e= xt4" in >> >> the body of a message to majordomo@vger.kernel.org >> >> More majordomo info at =A0http://vger.kernel.org/majordomo-info.h= tml >> > >> > >> > >> -- >> To unsubscribe from this list: send the line "unsubscribe linux-ext4= " in >> the body of a message to majordomo@vger.kernel.org >> More majordomo info at =A0http://vger.kernel.org/majordomo-info.html > > > -- To unsubscribe from this list: send the line "unsubscribe linux-ext4" i= n the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html