From mboxrd@z Thu Jan 1 00:00:00 1970 From: Fyodor Ustinov Subject: Re: rbd speed: write vs read Date: Fri, 06 May 2011 21:01:29 +0300 Message-ID: <4DC43779.3090506@ufm.su> References: <4DC40249.7020903@ufm.su> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit Return-path: Received: from mail.ufm.su ([77.120.103.19]:47461 "EHLO mail.ufm.su" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754470Ab1EFSBd (ORCPT ); Fri, 6 May 2011 14:01:33 -0400 In-Reply-To: Sender: ceph-devel-owner@vger.kernel.org List-ID: To: Sage Weil Cc: ceph-devel@vger.kernel.org On 05/06/2011 07:08 PM, Sage Weil wrote: > On Fri, 6 May 2011, Fyodor Ustinov wrote: >> Hi! >> >> oops. This my misconfiguration. > Out of curiousity, what was the misconfiguration? I have 20 OSD server. 10 in one vlan and 10 in other (btw, question: mon server can listen on all ip addressessimultaneously?). Because our NOC, 10 servers have anisotropic configuration - 800Mbps in one direction and 150Mbps to another. > Doing a streaming read in 10MB chunks when RBD is striped over 4MB > objects won't be terribly efficient. You should get the best sync read > performance when the object sizes match up. You'll get better performance > if whatever is sitting above the block layer dispatches multiple read > requests in parallel (i.e. something smarter than dd, like a file system). Hmm. But it's linear read. I do not know how it works internally, but did reading 10MB by 4MB chunk can not be executed in parallel? WBR, Fyodor.