From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from 87-104-106-3-dynamic-customer.profibernet.dk ([87.104.106.3]:47001 "EHLO kernel.dk" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753503Ab1ICD2q (ORCPT ); Fri, 2 Sep 2011 23:28:46 -0400 Message-ID: <4E619EE9.5030508@kernel.dk> Date: Fri, 02 Sep 2011 21:28:41 -0600 From: Jens Axboe MIME-Version: 1.0 Subject: Re: Bad performance when reads and writes on same LUN References: <116197A0179A3D4A95B135E5F0341FC11041B4DA@SACMVEXC2-PRD.hq.netapp.com> In-Reply-To: <116197A0179A3D4A95B135E5F0341FC11041B4DA@SACMVEXC2-PRD.hq.netapp.com> Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 7bit Sender: fio-owner@vger.kernel.org List-Id: fio@vger.kernel.org To: "Neto, Antonio Jose Rodrigues" Cc: zhu.yanhai@gmail.com, "Rettl, Matthias" , fio@vger.kernel.org On 2011-09-02 21:24, Neto, Antonio Jose Rodrigues wrote: > Thanks Jens, but why when we run READS only (not mixed) we don't have performance issues? (We run with direct=0)? Let me repeat: please don't top post! Etiquette on most open source lists are to bottom post, like you can see other people are doind. Apparently you don't have performance problems with buffered reads since you don't need a large queue depth to get good performance on reads alone. Since you seem a little lost, let me suggest that you start by diagnosing the raw read vs write performance of the device. Get rid of the file system, use filename=/dev/XXX directly to access that device. Beware that it will eat the data that is currently on the device. Once you get an idea for what the device can actually do, then you can start to consider what impact the file system has on that (if any, I looked at your job file, and each job would use separate files. so you should not have any buffered vs direct unhappiness going on). -- Jens Axboe