From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id ; Tue, 7 May 2002 12:26:52 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id ; Tue, 7 May 2002 12:26:51 -0400 Received: from gate.in-addr.de ([212.8.193.158]:25612 "HELO mx.in-addr.de") by vger.kernel.org with SMTP id ; Tue, 7 May 2002 12:26:50 -0400 Date: Tue, 7 May 2002 18:26:20 +0200 From: Lars Marowsky-Bree To: Jeff Dike Cc: linux-kernel@vger.kernel.org, user-mode-linux-devel@lists.sourceforge.net, user-mode-linux-user@lists.sourceforge.net Subject: Re: [uml-devel] Re: UML is now self-hosting! Message-ID: <20020507182620.U2539@marowsky-bree.de> In-Reply-To: <20020506181427.K918@marowsky-bree.de> <200205062055.PAA04067@ccure.karaya.com> Mime-Version: 1.0 Content-Type: text/plain; charset=iso-8859-1 Content-Disposition: inline Content-Transfer-Encoding: 8bit User-Agent: Mutt/1.3.22.1i X-Ctuhulu: HASTUR Sender: linux-kernel-owner@vger.kernel.org X-Mailing-List: linux-kernel@vger.kernel.org On 2002-05-06T15:55:52, Jeff Dike said: > > but spreading an instance across multiple nodes is nowhere as simple > > as it seems; > It is if you want to be sufficiently stupid about it :-) Ugh. > > where do you keep OS data, > It gets faulted from host to host as needed. Ugh Ugh Ugh. You need coherency algorithms for these. > > IO access, > This scheme (and any clustering scheme, I think) would need back channels > for one node to access the devices of another Right. You need communication services and coherency algorithms for these ;-) > > scheduling decisions, > This machine thinks it's a normal SMP box, so scheduling happens as normal Ugh ugh ugh. Too many page faults; you need a scheduler capable of keeping node affinity. > > inter-node communication in the first place, how to deal > > with node failure etc... > Maybe I'm not familiar enough with the clustering world, but I was under the > impression that with a normal SSI cluster, the nodes are like CPUs in an > SMP box - if one fails, the whole thing dies. In other words, that SSI > clustering and HA clustering are pretty disjoint. No. In the optimal case, nothing dies. ;-) In the less than optimal case, only the processes affected directly by the failure die (the processes which had dirty pages there etc). In the really useless case, the entire cluster goes down like on a SMP box or an unpartitioned CC-NUMA. "In search of clusters" is definetely highly recommended reading. It is very entertaining, too. Sincerely, Lars Marowsky-Brée -- Immortality is an adequate definition of high availability for me. --- Gregory F. Pfister