Newsgroups: comp.parallel
From: ntuck@muddcs.cs.hmc.edu (Nathan D. Tuck)
Subject: Re: SCI Out of Gas at Starting Gate?
Keywords: SCI
Organization: Harvey Mudd College, Claremont, CA
Date: Thu, 30 Nov 1995 19:24:27 GMT
Message-ID: <49l0db$gpu@jaws.cs.hmc.edu>

In article <49fk7s$mtc@sdcc12.ucsd.edu>,
Keith Muller <muller@sdcc33.ucsd.edu> wrote:
>
>While doing some very rough performance estimates the other day with a grad 
>student about using SCI rings to create a CC-NUMA SMP some very odd results
>came up.  Does anyone see what is wrong in either assumptions or calculations?

>o  12 Million * .5 = 6M cache cycles/sec/quad PentiumPro on SCI
>               ^^^ probability of cache line being on local node

I'm going to say that this is the assumption that is really killing
you.  Any big NUMA machine is going to have to handle process
migration in an intelligent manner.  In addition, there is likely to
be a lot of page duplication throughout the system, particularly of
read only pages.  This is going to cut down on SCI transactions by a
significant factor (at least if these are ever going to work).  Your
model would require assuming that you have all the processors working
on the same data set and doing a lot of reads and writes on the same
pages or cache lines.  That'll kill most any parallel system, even an
SMP.  

Of course this doesn't mean that doing a NUMA using SCI is going to be
easy...

nate



