Path: utzoo!attcan!uunet!snorkelwacker!ai-lab!teenage-mutant!misha
From: misha@teenage-mutant.ai.mit.edu (Mike Bolotski)
Newsgroups: comp.arch
Subject: Re: Shared Memory (was Terradata architecures)
Message-ID: <11182@life.ai.mit.edu>
Date: 4 Oct 90 16:57:18 GMT
References: <1990Oct1.200613.635@tera.com> <1990Oct3.013509.1470@news.iastate.edu> <10651@pt.cs.cmu.edu>
Sender: news@ai.mit.edu
Organization: MIT Artificial Intelligence Laboratory
Lines: 53

In article <10651@pt.cs.cmu.edu> lindsay@gandalf.cs.cmu.edu (Donald Lindsay) writes:
>It is clear that it is simplest to design a message-based ensemble
>machine.  A lot of issues can be punted to software, and the hardware
>that is designed, pretty well has to be there in any case.
>
>It is also clear that a shared-memory ensemble machine is desirable.

This is not necessarily clear.  There are various penalties
associated with shared memories, including performance and cost.

>Sharing allows some nice programming techniques, and makes a
>difference when trying to port a lot of software to the new machine.

Granted, most parallel software was written under the shared memory
model, and if the main constraint is porting existing software, then
shared memory machines win. This argument was used in the past
on various innovations.

>Next, at what level will we share?  If it's at the level of virtual
>memory pages, then this is already being done, over LANs, by several
>operating systems (e.g. Mach).  The only hardware facilities required
>are 1) a decent MMU on each node, and 2) page-sized messages.

3) a tolerable interconnection network.

While sharing at the page level across a network works, increasing
the number of workstations to say, 1000, will seriously impede
performance.

>If a machine is to have finer-grained sharing, then a number of
>efficiency issues come up.  I am aware of a variety of approaches,
>each with its own penalty or price.  I believe that eventually, one
>of these will be implemented in a way that adds essentially nothing
>to the machine's manufacturing cost.  

I disagree.  Shared memory automatically creates contention at
for frequently used data.  Caching is a possible solution, but 
cache coherence protocols for fine grained machines become incredibly
complex, and further impact performance.

Further, shared memory does not obey the natural laws of physics,
in a sense.  The abstract shared memory model pretends that there
is a large number of communication paths to a single point. This
simply isn't true, and additional hardware is required to provide
this illusion to the programmer. This hardware costs time and
dollars.

Parallel programs designed to reflect the underlying physical
reality will operate faster than those that work in a virtual
model. I suppose an IMHO is in order here.

Mike Bolotski          Artificial Intelligence Laboratory, MIT
misha@ai.mit.edu       Cambridge, MA