[Beowulf] Network considerations for new generation cheap beowulfcluster
Peter St. John
peter.st.john at gmail.com
Thu May 24 09:24:58 PDT 2007
Well I remember when people were talking about distributed computing and
they said, "well, this would be trivial if you could just boot unix on all
the nodes, haha". So I'm still drawn to the idea of something small, like
embedded ROMDOS, on a compute node. But RAM is free and IBM put linux on a
wristwatch (honkin' big power suplly, by the standards of wrist watches, but
it ran X :-)
On 5/24/07, Larry Stewart <larry.stewart at sicortex.com> wrote:
> Peter St. John wrote:
> > This led me to SUNMOS (OS for parallel processing, Sandia's
> > alternative for the aforementioned intel
> > PSC descendants) http://en.wikipedia.org/wiki/SUNMOS. Sandia's page
> > looks like real progress for inertial containment sustainable
> > thermonuclear fusion, which really wouuld be super duper cool, fuel
> > your zeppelin with seawater, but I can't find latter-day references to
> > SUNMOS. Anybody know what became of it
> I don't know what happened to SUNMOS, but the intellectual descendents
> of it are the microkernels like L4
> and particularly the FASTOS project, see http://www.cs.unm.edu/~fastos/
> There's a group who think that full function OSs like Linux are
> unnecessary or wasteful for clusters. The arguments
> include virtual memory being unneeded or slow and OS activity ("OS
> noise") limiting scaling of applications.
> As a consequence, you see things like BlueGene/L with a microkernel on
> the compute nodes and Linux on the I/O nodes.
> I don't necessarily believe the arguments, but I like tinkering with a
> new tricked out OS as well
> as the next guy.
> Just for fun, I tracked down who said "When you hear 'virtual', you
> should think 'slow'." -- it was Dave Clark of MIT.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Beowulf