Path: utzoo!attcan!uunet!yale!lisper-bjorn
From: lisper-bjorn@CS.YALE.EDU (Bjorn Lisper)
Newsgroups: comp.arch
Subject: Re: Is Shared Memory Necessary?
Message-ID: <29577@yale-celray.yale.UUCP>
Date: 18 May 88 21:55:46 GMT
References: <503@xios.XIOS.UUCP> <2676@pdn.UUCP> <674@cernvax.UUCP> <685@thalia.rice.edu> <8125@pur-ee.UUCP>
Sender: root@yale.UUCP
Reply-To: lisper-bjorn@CS.YALE.EDU (Bjorn Lisper)
Organization: Yale University Computer Science Dept, New Haven CT  06520-2158
Lines: 36

In article <8125@pur-ee.UUCP> hankd@pur-ee.UUCP (Hank Dietz) gives a list of
MIMD machines with large-scale shared memory address space but physically
distributed memory:
....
>CHoPP
>Denelcor HEP
>NYU Ultra
>BBN Butterfly
>IBM RP3
>RFM-MIMD
>Cedar
>BBM Monarch
>CARP Machine
>Horizon/Tera
....

Is the list restricted to existing machines? If not, then I wold like to add
the "Fluent Machine" being projected here at Yale. Physically this consists
of processors with local memory connected in a butterfly pattern. These
local memories will however be seen as one shared memory by programs. To
accomplish this messages have to be routed in the network and this is done
by something called "fluent routing", which is the essential new thing for
this machine. It will also support combining operations such as full
multiprefix.

If I am to give my personal opinion on very large scale global address space
machines, then it is that I think they must provide local, fixed routing
between close processors as an alternative addressing mode. The generality
of the global address scheme will always have to be paid in long access time
compared with local access.  Many algorithms, especially in fields such as
scientific computing, have a quite data-independent structure which means
that the communication pattern between the computation steps can be
determined at compile-time and mapped to the network of processors to
provide a high degree of locality.

Bjorn Lisper