From mboxrd@z Thu Jan 1 00:00:00 1970 From: David Mosberger Date: Wed, 06 Apr 2005 00:23:51 +0000 Subject: Re: [PATCH] add a clear_pages function to clear pages of higher Message-Id: <16979.11287.36091.610287@napali.hpl.hp.com> List-Id: References: <200503111008.12134.vda@port.imtp.ilyichevsk.odessa.ua> <200503181154.37414.vda@port.imtp.ilyichevsk.odessa.ua> <20050318192808.GB38053@muc.de> <16963.2075.713737.485070@napali.hpl.hp.com> In-Reply-To: MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: Christoph Lameter Cc: davidm@hpl.hp.com, Andi Kleen , Christoph Lameter , Denis Vlasenko , Dave Hansen , Andrew Morton , Linux Kernel Mailing List , Mel Gorman , linux-ia64@vger.kernel.org, Jens.Maurer@gmx.net >>>>> On Tue, 5 Apr 2005 17:15:53 -0700 (PDT), Christoph Lameter said: Christoph> On Thu, 24 Mar 2005, David Mosberger wrote: >> That's definitely the case. See my earlier post on this topic: >> http://www.gelato.unsw.edu.au/linux-ia64/0409/11012.html >> Unfortunately, nobody reported any results for larger machines >> and/or more interesting workloads, so the patch is in limbo at >> this time. Clearly, if the CPU that's clearing the page is >> likely to use that same page soon after, it'd be useful to use >> temporal stores. Christoph> Here are some numbers using lmbench of temporal writes Christoph> vs. non temporal writes on ia64 (8p machine but lmbench Christoph> run only for one load). There seems to be some benefit Christoph> for fork/exec but overall this does not seem to be a Christoph> clear win. I suspect that the distinction between Christoph> temporal vs. nontemporal writes is be more beneficial on Christoph> machines with smaller pagesizes since the likelyhood that Christoph> most cachelines of a page are used soon is increased and Christoph> therefore hot zeroing is more beneficial. What LMbench test other than fork/exec would you have expected to be affected by this? LMbench is not a good benchmark for this (remember: it's a _micro_ benchmark). --david