Newsgroups: comp.parallel.mpi
From: I.J.Bush@dl.ac.uk (Ian the Shrublet)
Subject: Re: MPI vs native-CRAY
Organization: Daresbury Lab, Warrington, U.K.
Date: 5 Sep 1996 13:15:57 GMT
Message-ID: <50mjqd$eaf@mserv1.dl.ac.uk>

In article <322DE865.167E@nas.nasa.gov>, Stephen Barnard <barnard@nas.nasa.gov> writes:
|> Ian the Shrublet wrote:
|> > 
|> > In article <322DA5B6.2EAC@megafauna.com>, Stephen Barnard <steve@megafauna.com> writes:
|> 

SNIP a few technical details and an amicable disagreement ( is having
an amicable disagreement netiquette ? Surely the flamethrowers should
be out ? :) )

|> 
|> Your experience must be different from mine.  I used to work for Cray,
|> and I used the shmem stuff intensively from day 1 of the T3D, even
|> before it was called shmem.  (There's an interesting and funny story
|> about how shmem came to be, but it's a little delicate so I can't go
|> into it.)
|> 
|> I never found a single bug.  A couple of times I thought I had, but I
|> was wrong.  I used it for some *extremely* fancy programs, if I do say
|> so myself.  :-)  It always worked like a charm, as long as you're
|> careful about cache coherence with puts, or you use shmem_set_cache_inv,
|> which is the best thing to do.
|> 
|> The shmem stuff is just a low-level interface to the logic that Cray
|> added to support distributed shared memory.  There's really not much
|> there to go wrong.  As long as the hardware works the shmem stuff is
|> likely to work.

In fact the experiences arent all that different. shmem, once you
get used to it, is very straightforward to use and you can do
some very snazzy and efficient stuff with it. In fact it's probably
easier than getting decent single node performance. However it's the
getting used to it bit that is the problem, IMHO, for new users. If
Johnny Newuser does not know about the issues involved, especially
wrt the cache, he can come unstuck.

|> 
|> The only disadvantage to shmem, IMHO, is that it isn't portable.  It's a
|> shame that people have to use MPI on the T3D solely for portability,
|> because that doesn't take best advantage of the really outstanding
|> feature of this machine -- it's low-latency, lightweight communication.
|> 

True. Let's hope that one sided comms are in MPI-2 and in a form that is
readily compatible with shmem.


|> BTW, there is an even better way to use the T3D's IPC hardware that
|> avoids the procedure-call overhead of shmem, but I don't believe that
|> Cray supports it.  If anyone is interested they can send me email and
|> I'll put them in touch with the right person at Cray/SGI.
|> 


I'd be interested, if only for curiosity !

|> What is a Shrublet?  A little shrub?

Sort of. Look at my surname in the reply-to field. Then consider that
I'm 5'4''.



Ian

p.s. ( another one ! ) Which part of Cray did you work for ? I know,
well have occassional contact with, some Cray employees.





