From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mx1.redhat.com ([209.132.183.28]:58073 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753872Ab0DWW0r (ORCPT ); Fri, 23 Apr 2010 18:26:47 -0400 Message-ID: <4BD21EA2.9040502@RedHat.com> Date: Fri, 23 Apr 2010 18:26:42 -0400 From: Steve Dickson To: Trond Myklebust CC: Linux NFS Mailing list Subject: Re: NFS client send out of order offsets.... References: <4BD21777.1060305@RedHat.com> <1272060583.3019.11.camel@localhost.localdomain> In-Reply-To: <1272060583.3019.11.camel@localhost.localdomain> Content-Type: text/plain; charset=UTF-8 Sender: linux-nfs-owner@vger.kernel.org List-ID: MIME-Version: 1.0 On 04/23/2010 06:09 PM, Trond Myklebust wrote: > On Fri, 2010-04-23 at 17:56 -0400, Steve Dickson wrote: >> Hey Trond (and all), >> >> It been brought to my attention that Linux NFS clients, with big-ish >> writes (8GB and above), quickly started sending out writes with >> out-of-order offsets... Meaning in the start of an 8GB write a 7GB offset >> will be send and then smaller offsets, closer to the the beginning >> of the file, will follow. >> >> Now I realize this is perfectly fine from a protocol standpoint and (w/out >> any on hands investigation) pretty sure it has to do with how pages are >> being kicked out of the cache (i.e. memory pressure), meaning is >> not an NFS issue at all... but... >> >> It was also point out to me that this type of out of order-ness does >> not happen with Solaris clients and with an AIX clients there is some >> mount option fix out of order-ness... >> >> So my question is there some way to tweak the client to ensure >> offset are send sequential order... >> >> Note this happening on a older kernel, so before I go off and >> "re-invent the wheel", I was wonder if other people have seen this >> issue and has it been addressed... >> >> tia, >> >> steved. > > The NFS layer will always write out pages in order whenever it initiates > a flush. > > The problem here is the VM's use of the 'range_cyclic' writeback control > option in functions like 'balance_dirty_pages()' and in the kflush > daemons. > That neither can nor should be fixed in the NFS layer. It really needs > to be addressed in the VM. > Understood... completely.... But I wonder if there some vm tweak (via sysctl) that would give a better chance of having things go out in order... steved.