Path: utzoo!utgpu!jarvis.csri.toronto.edu!rutgers!tut.cis.ohio-state.edu!ucbvax!decwrl!elroy.jpl.nasa.gov!cit-vax!jerbil
From: jerbil@cit-vax.Caltech.Edu (Joseph Beckenbach)
Newsgroups: comp.os.mach
Subject: Re: MACH on Intel-Hypercube
Message-ID: <14194@cit-vax.Caltech.Edu>
Date: 8 Mar 90 22:27:37 GMT
References: <9902@spool.cs.wisc.edu>
Reply-To: jerbil@csvax.caltech.edu (Joseph Beckenbach)
Distribution: usa
Organization: CS department, Caltech, Pasadena, California
Lines: 23

In his article Manolis Tsangaris <mt@cleo.cs.wisc.edu> writes:
>Since IPSC/2 is a "shared nothing" multiprocessor, is it advisable to run a
>single MACH kernel over its 32 (or less) processors?  The alternative solution
>would be to treat hypercube as a super fast distributed system, with one MACH
>kernel per processor. But this would make expensive things like task migration
>(how about TLB consistency!).

	The only two sensible things to seem to be to place one instance
of the MACH kernel on each node and act as if they were on a single private
network, or to combine several nodes with some form of shared memory emulator
(an external pager?) and work from there.  For anything smaller than 4MB I
can see no alternatives -- someone care to enlighten us further?

	[I had enquired about hypercubes running MACH when I attended the
MACH tutorials at the Winter 1990 USENIX;  Tevanian and Nawar both seemed to
think that placing MACH on hypercubes would be an interesting exercise and
little more -- I may have misinterpreted.]

		Joe Beckenbach

-- 
Joe Beckenbach		 jerbil@csvax.caltech.edu	VEGGIES FOREVER!
	Toto, kansasoseum nun est cognito.  -- Farley