From owner-freebsd-current Tue Dec 29 20:50:03 1998 Return-Path: Received: (from majordom@localhost) by hub.freebsd.org (8.8.8/8.8.8) id UAA29817 for freebsd-current-outgoing; Tue, 29 Dec 1998 20:50:03 -0800 (PST) (envelope-from owner-freebsd-current@FreeBSD.ORG) Received: from ns1.adsu.bellsouth.com (ns1.adsu.bellsouth.com [205.152.173.2]) by hub.freebsd.org (8.8.8/8.8.8) with ESMTP id UAA29748 for ; Tue, 29 Dec 1998 20:50:00 -0800 (PST) (envelope-from ck@ns1.adsu.bellsouth.com) Received: (from ck@localhost) by ns1.adsu.bellsouth.com (8.9.1a/8.9.1) id XAA06549; Tue, 29 Dec 1998 23:49:37 -0500 (EST) Message-ID: <19981229234937.N477@ns1.adsu.bellsouth.com> Date: Tue, 29 Dec 1998 23:49:37 -0500 From: Christian Kuhtz To: "Steven P. Donegan" , Steve Passe Cc: freebsd-current@FreeBSD.ORG Subject: Re: NOW/MOSIX/Beowulf References: <199812300239.TAA29911@Ilsa.StevesCafe.com> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii X-Mailer: Mutt 0.93.2i In-Reply-To: ; from Steven P. Donegan on Tue, Dec 29, 1998 at 06:46:01PM -0800 Sender: owner-freebsd-current@FreeBSD.ORG Precedence: bulk X-Loop: FreeBSD.ORG > 1) SMP base Actually, it wouldn't really matter whether the node is SMP or not. You're talking about building a distributed memory multiprocessor and a design goal for the cluster should be that it is irrelevant whether SMP is there or not. Make it machine architecture independant. Who knows, maybe somebody will port FreeBSD to a distributed memory multiprocessor and then what? > 2) NOW aware (MOSIX migration under load model is good) Hmm. Yeah, certainly would be nice. > 3) The network is king - so Cisco or Beowulf style etherchannel or > gigabit or gigabit etherchannel is good. .. do we have GE supported cards in the code base today? > 4) Network switching fabric - HP and Cisco both support etherchannel - > 400 megabits full duplex is a decent backbone :-) 4 gigs in Cisco only > environs but I doubt any PCI card can do much with that :-) Why bother doing EtherChannel? You need special PCI cards and all that. It would make much more sense to use GE. Money no object, Catalyst 550x's work great. We got the GE stuff in the lab. You also need to think about what your overall node architecture looks like. Do you want to build a hypercube perhaps? An any to any matrix over an Ethernet fabric? > 5) Storage - Network Appliances support gig ethernet - and can actually > serve that bandwidth :-) Why not use FreeBSD as a storage subsystem? NetApps aren't exactly cheap. A real supercomputer wannabe needs an HMS, too. And another item to your list should be 6) APIs What good is this if you can't write to it? PVM comes to mind (or its successor), and MPI. I think I (and many others) would be happy if they had a system similiar to an RS/6000 SP2, which is essentially the same as what you're talking about. And I don't need the OS to do thread migration (which particularly in a distributed memory model is expensive). It depends very much on what you want this box to do. Most supercomputer architectures (and clusters) are designed for particular application characteristics. You need to think about your scope. Are you trying to solve the avid hacker's cluster dream with a hodge podge array of machinery or are you building a cluster consisting of n # of nodes doing parallel processing, which each node being pretty much a standard config.. Trying to be the be all end all usually isn't a good idea for massively parallel situations. My $.02, Chris -- Frisbeetarianism, n.: The belief that when you die, your soul goes up on the roof and gets stuck. To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-current" in the body of the message