From mboxrd@z Thu Jan 1 00:00:00 1970 From: Andrea Arcangeli Subject: Re: Benchmarking on CentOS 5 Date: Sun, 1 Jun 2008 16:32:59 +0200 Message-ID: <20080601143259.GE28289@duo.random> References: <483D9AE5.6040509@cs.ualberta.ca> <20080528183352.GE8086@duo.random> <483DC01C.5090603@bppiac.hu> <20080529092309.GH8086@duo.random> <483E81AD.9050209@bppiac.hu> <483F75CC.9010303@cisco.com> <48400371.3060208@cisco.com> <484039C1.2010806@bppiac.hu> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: "David S. Ahern" , kvm@vger.kernel.org To: Farkas Levente Return-path: Received: from host36-195-149-62.serverdedicati.aruba.it ([62.149.195.36]:44451 "EHLO mx.cpushare.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751859AbYFAOdD (ORCPT ); Sun, 1 Jun 2008 10:33:03 -0400 Content-Disposition: inline In-Reply-To: <484039C1.2010806@bppiac.hu> Sender: kvm-owner@vger.kernel.org List-ID: On Fri, May 30, 2008 at 07:30:41PM +0200, Farkas Levente wrote: > without crash) and the performance is similar (ie 80-90%) to the > development version. unfortunately it seems currently there is no such I hope you didn't reach this conclusion about different performance between enterprise kernel host and mainline kernel host because I didn't answer promptly to your previous email, sorry! With regard to your performance question, preempt notifier emulation really shouldn't be noticeable in real life. It worth to run the fastest code only if you're doing pure benchmarking, in which case you want to avoid any unnecessary exception and heavyweight exit (that should reduces the time to schedule of a couple thousand cycles I guess, but re-schedule events aren't so frequent and this is a per-task per-cpu breakpoint with my latest emulation logic taking advantage of the __switch_to internals). I can hardly measure any difference here between SLES10 SP2 host and mainline host with a simple dd from pagecache to /dev/null. The change between enterprise kernel and current mainline kernel may be more significant for the final performance than the impact of the preempt notifier emulation. In any case the difference should be much less than 5% as far as preempt notifier emulation is concerned. It should be couple usec lost every couple msec or so.