From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S932442AbXBNS5O (ORCPT ); Wed, 14 Feb 2007 13:57:14 -0500 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S932447AbXBNS5O (ORCPT ); Wed, 14 Feb 2007 13:57:14 -0500 Received: from omx1-ext.sgi.com ([192.48.179.11]:43322 "EHLO omx1.sgi.com" rhost-flags-OK-OK-OK-FAIL) by vger.kernel.org with ESMTP id S932442AbXBNS5N (ORCPT ); Wed, 14 Feb 2007 13:57:13 -0500 Date: Wed, 14 Feb 2007 10:57:00 -0800 (PST) From: Christoph Lameter To: Nick Piggin cc: Linux Kernel Mailing List , Linux Memory Management List Subject: Re: [patch] mm: NUMA replicated pagecache In-Reply-To: <20070213060924.GB20644@wotan.suse.de> Message-ID: References: <20070213060924.GB20644@wotan.suse.de> MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org On Tue, 13 Feb 2007, Nick Piggin wrote: > Just tinkering around with this and got something working, so I'll see > if anyone else wants to try it. > > Not proposing for inclusion, but I'd be interested in comments or results. We would be very interested in such a feature. We have another hack that shows up to 40% performance improvements. > At the moment the code is a bit ugly, but it won't take much to make it a > completely standalone ~400 line module with just a handful of hooks into > the core mm. So if anyone really wants it, it could be quite realistic to > get into an includable form. Would be great but I am a bit skeptical regarding the locking and the additonal overhead moving back and forth between replications and non replicated page state. > At some point I did take a look at Dave Hansen's page replication patch for > ideas, but didn't get far because he was doing a per-inode scheme and I was > doing per-page. No judgments on which approach is better, but I feel this > per-page patch is quite neat. Definitely looks better. > - Would be nice to transfer master on reclaim. This should be quite easy, > must transfer relevant flags, and only if !PagePrivate (which reclaim > takes care of). Transfer master? Meaning you need to remove the replicated pages? Removing of replicated pages should transfer reference bit? > - Should go nicely with lockless pagecache, but haven't merged them yet. When is that going to happen? Soon I hope?