Date: Fri, 19 Jan 2001 09:21:59 -0700 (MST) From: Ronald G Minnich <rminnich@lanl.gov> To: Brooks Davis <brooks@one-eyed-alien.net> Cc: "Russell L. Carter" <rcarter@pinyon.org>, Wes Peters <wes@softweyr.com>, Uwe Pierau <uwe.pierau@tu-clausthal.de>, <freebsd-hackers@FreeBSD.ORG> Subject: Re: Clustering FreeBSD Message-ID: <Pine.LNX.4.30.0101190910030.26378-100000@mini.acl.lanl.gov> In-Reply-To: <20010119073826.A30053@Odin.AC.HMC.Edu>
next in thread | previous in thread | raw e-mail | index | archive | help
On Fri, 19 Jan 2001, Brooks Davis wrote: > For those who want a simple, stupid way to do this, making an MPI > application is a convenient first step. MPI is pretty similar to PVM > except that I don't know of anyone in the high performance computing > community that still uses PVM for new applications (I'm sure they exist, > but they are not exactly common.) For some reason the Open Source > community still has this bizare idea that PVM is the way to go. MPI is way to heavy for process spawning. We have measured appallingly long times here on our clusters, up to 30 seconds just to get things running on 64 nodes. I keep offering this, and keep getting no takers, but I do have a tool called vex that will get 128 processes running on 128 nodes in 1/2 second. See it at http://www.lanl.gov/~rminnich/. That's the level of performance you want for a start. It actually runs tons better on FreeBSD than on Linux due to Linux TCP silliness. You really want a single login, single IP address, cluster. There's an example: http://www.scyld.com. You need a process model that's much more capable than what we have now. Three ways to go that I can think of. The worst is to nfs-mount all the /proc on the front-end. Yuck. The second-coolest-thing to do is to build a "bproc"-like interface for freebsd. The absolute coolest thing is (do I repeat myself :-) put plan-9 style remote process and private name spaces into freebsd. Before you comment on the plan9 idea, if you have not read the docs, go read them. Bproc gives you a global name space for processes, which is ok but not as scalable as the plan9 approach. But the Scyld stuff is quite nice, we're using it here on two clusters. ron To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-hackers" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?Pine.LNX.4.30.0101190910030.26378-100000>