From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1757338AbaCRR0q (ORCPT ); Tue, 18 Mar 2014 13:26:46 -0400 Received: from qmta12.emeryville.ca.mail.comcast.net ([76.96.27.227]:49598 "EHLO qmta12.emeryville.ca.mail.comcast.net" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1757151AbaCRR0m (ORCPT ); Tue, 18 Mar 2014 13:26:42 -0400 Date: Tue, 18 Mar 2014 12:26:40 -0500 (CDT) From: Christoph Lameter X-X-Sender: cl@nuc To: Peter Zijlstra cc: Dave Hansen , lsf@lists.linux-foundation.org, Linux-MM , Wu Fengguang , LKML Subject: Re: [Lsf] [LSF/MM TOPIC] Testing Large-Memory Hardware In-Reply-To: <20140318165059.GI22095@laptop.programming.kicks-ass.net> Message-ID: References: <5328753B.2050107@intel.com> <20140318165059.GI22095@laptop.programming.kicks-ass.net> Content-Type: TEXT/PLAIN; charset=US-ASCII Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Tue, 18 Mar 2014, Peter Zijlstra wrote: > > My gut reaction was that we'd probably be better served by putting > > resources in to systems with higher core counts rather than lots of RAM. > > I have encountered the occasional boot bug on my 1TB system, but it's > > far from a frequent occurrence, and even more infrequent to encounter > > things at runtime. > > > > Would folks agree with that? What kinds of tests, benchmarks, stress > > tests, etc... do folks run that are both valuable and can only be run on > > a system with a large amount of actual RAM? > > We had a sched-numa + kvm fail on really large systems the other day, > but yeah in general such problems tend to be rare. Then again, without > test coverage they will always be rare, for even if there were problems, > nobody would notice :-) SGI had systems out there up to few PB of RAM. There were a couple of tricks to get this going. Bootup time was pretty long. I/O has to be done carefully. The MM subsystem used to work with these sizes (I have not had a chance to verify that recently). This was Itanium with 64K page size so you had a factor of 16 less page structs to process. What I saw there is one of the reasons why I would like to see larger page support in the kernel. Managing massive amounts of 4k pages is creation far too much overhead.