Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!cwjcc!gatech!hubcap!blevins
From: blevins@mcnc.org (Donald W. Blevins)
Newsgroups: comp.parallel
Subject: performance question
Message-ID: <4361@hubcap.UUCP>
Date: 9 Feb 89 15:43:18 GMT
Sender: fpst@hubcap.UUCP
Lines: 20
Approved: parallel@hubcap.clemson.edu

I would like to hear opinions on the use of feature that we
have added to our own SIMD machine.  As a background, it
is based on the NASA MPP, but with some important new changes.
The machine is a SIMD, 2D mesh (includes diagonals).  Each
PE is bit-serial, and has access to 1k bits of static ram locally.

We have the capability to locally modify the SIMD global address,
and thus generate an offset unique to each PE.

We know of several applications that can use this feature, but
we are unable to quantify the performance improvement.

Do you know of an algorithm(s) that can make use of this feature,
and what the performance improvement is over the same machine
without this feature?

Thanks.

Don Blevins
MCNC