From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1754274AbZBBIxv (ORCPT ); Mon, 2 Feb 2009 03:53:51 -0500 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1751219AbZBBIxn (ORCPT ); Mon, 2 Feb 2009 03:53:43 -0500 Received: from bombadil.infradead.org ([18.85.46.34]:54869 "EHLO bombadil.infradead.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750780AbZBBIxm (ORCPT ); Mon, 2 Feb 2009 03:53:42 -0500 Subject: Re: hackbench [pthread mode] regression with 2.6.29-rc3 From: Peter Zijlstra To: "Zhang, Yanmin" Cc: Lin Ming , linux-kernel In-Reply-To: <1233537134.2604.24.camel@ymzhang> References: <1233473426.2604.13.camel@ymzhang> <1233476961.13659.12.camel@minggr.sh.intel.com> <1233479836.4787.63.camel@laptop> <1233482239.4787.65.camel@laptop> <1233537134.2604.24.camel@ymzhang> Content-Type: text/plain Date: Mon, 02 Feb 2009 09:53:38 +0100 Message-Id: <1233564818.4787.107.camel@laptop> Mime-Version: 1.0 X-Mailer: Evolution 2.24.3 Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Mon, 2009-02-02 at 09:12 +0800, Zhang, Yanmin wrote: > process timer (by setitimer) isn't good. Is per-cpu itimer to improve it? > So the per-cpu itimer could improve this situation when thread number is far bigger than > cpu number. I didn't retry specweb2005 with 2.6.28. 1) process wide itimers are rubbish, 2) per-cpu itimers are rubbish too, for the very simple reason they waste gobs of memory for sane programs. I would rather go back to the old model where we iterate all threads, and find a way to not make programs with too many threads for their own good lock up the kernel, but instead get poor service. Now the problems appears to be that I overlooked that we keep the itimer clock running at all times, not only when we have a pending timer, and I suppose that the standard mandates that behaviour :/ Anyway, I will try to sort something out, at worst we'll have to revert to the .28 state and live with the per-cpu crap for another release.