From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S266692AbUHOOOO (ORCPT ); Sun, 15 Aug 2004 10:14:14 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S266701AbUHOOOO (ORCPT ); Sun, 15 Aug 2004 10:14:14 -0400 Received: from dbl.q-ag.de ([213.172.117.3]:52368 "EHLO dbl.q-ag.de") by vger.kernel.org with ESMTP id S266692AbUHOOOK (ORCPT ); Sun, 15 Aug 2004 10:14:10 -0400 Message-ID: <411F7067.8040305@colorfullife.com> Date: Sun, 15 Aug 2004 16:17:11 +0200 From: Manfred Spraul User-Agent: Mozilla/5.0 (X11; U; Linux i686; fr-FR; rv:1.7.2) Gecko/20040803 X-Accept-Language: en-us, en MIME-Version: 1.0 To: Christoph Lameter CC: linux-kernel@vger.kernel.org Subject: Re: page fault fastpath: Increasing SMP scalability by introducing pte Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org Christoph wrote: >Well this is more an idea than a real patch yet. The page_table_lock >becomes a bottleneck if more than 4 CPUs are rapidly allocating and using >memory. "pft" is a program that measures the performance of page faults on >SMP system. It allocates memory simultaneously in multiple threads thereby >causing lots of page faults for anonymous pages. > > > Very odd. Why do you see a problem with the page_table_lock but no problem from the mmap semaphore? The page fault codepath acquires both. How often is the page table lock acquired per page fault? Just once or multiple spin_lock calls per page fault? Is the problem contention or cache line trashing? Do you have profile/lockmeter output? Is the down_read() in do_page_fault() a hot spot, too? -- Manfred