Newsgroups: comp.parallel
From: strauss@convex.convex.com (Henry Strauss)
Subject: Re: KSR reliability (was Re: SMP vs. MPP)
Organization: CONVEX Computer Corporation
Date: Tue, 10 Jan 1995 13:23:07 GMT
Message-ID: <3erc38$f6@convex.convex.com>

In article <ACAIRD.95Jan6113351@meson.engin.umich.edu>, acaird@engin.umich.edu (Andrew Justin Caird) writes:
|> In article <dm-0601950219560001@128.89.19.84> dm@bbn.com (David Mankins) writes:
|>    From: dm@bbn.com (David Mankins)
|>    Date: Fri, 6 Jan 1995 15:56:23 GMT
|> 
|>    In article <3eb530$6bk@convex.convex.com>, strauss@convex.convex.com
|>    (Henry Strauss) wrote:
|> 
|>    > 
|>    > Yes, there was a reliability problem -- in fact more than one --
|>    > but I don't see this as anything surprising.  The KSR machines were
|>    > still relatively new products, and AllCache the first hardware
|>    > implementation of virtually shared memory.
|> 
|>    Surely the BBN GP1000 (a later version of the Butterfly) had
|>    virtually shared memory --- they ran Mach.  On the other hand, a lot of
|>    BBN Butterfly hardware engineers moved to KSR after BBN announced
|>    it was getting out of the parallel processor business.
|> 
|>    - dave mankins (dm@bbn.com, dm@world.std.com)
|> 
|> Indeed it was a very similar arch. 
|> 
I agree and disagree.
After reading a bit more on the Butterfly and its successors [in:
Past, Present, Parallel by Trew & Wilson], I agree that it also was
a parallel machine with physically distributed but logically shared
memory (with NUMA).
However, I wouldn't say that the _architectures_ are very similar.
The systems were similar in the way they were programmed (namely
the shared memory paradigm).  But from the architectural point of
view the KSR is unique because of its memory system: AllCache means
that all physical memory in the machine is cache, and an address
has NO fixed location within the system.
-Henry



