DEPARTMENT OF MATHEMATICS AND COMPUTER SCIENCE               
                    UNIVERSITY OF SOUTHERN DENMARK, ODENSE                   

                          Distributed Shared Memory                          

                                 Brian Vinter                                
                              Tromsoe University                             

                     Tuesday, March 16, 1999, at 2:15 PM                     
                               The Seminar Room                              


Conventional wisdom on  how to build  and program supercomputers  is rapidly
disintegrating.  Machine architecture and programming  paradigm are becoming
disassociated and portability  is becoming  a topic  even at  this level  of
computation. While nobody can safely say to know what lies ahead in the area
of  supercomputing  certain  features  are  becoming  clearer  on  tomorrows
supercomputers.  The vector processor is  merging with conventional  CPU and
massively parallel processors  are starting  to support  shared memory  as a
programming paradigm.

Perhaps the most  exiting aspect of  this transition is  that supercomputers
are changing from large  custom systems into more  commodity parts.  This in
essence means that you can  build your own supercomputer  from parts you buy
in your local  PC store.  In  fact, for  Kr.  500K one  can build  a 64  CPU
supercomputer with a sustained performance close to 20 Gflop/second.  Enough
to join the list of the worlds top 500 fastest computers.

The problem that  arises is  then how to  program these  new supercomputers.
Programming with  explicit  message  passing is  inherently  hard,  and  any
application requires re-orchestration for  a new architecture or  even a new
configuration of an existing architecture. From this point of view, a shared
memory  programming  paradigm  preferable.  Unfortunately,  physical  shared
memory does not  scale well, and  anything above  32 CPUs yield  rather poor
performance.  The solution  it  turns out  is  to  provide a  shared  memory
programming API on a distributed memory architecture.

This talk will  introduce and describe  the known approaches  to distributed
shared memory, DSM. We will then go into the details of one such system, the
PastSet model, which  we are currently  working on in  Tromsoe.  The PastSet
programming model  will be  described and  it will  be  shown how  a set  of
applications  that  use  the  current  PastSet  implementation  is  able  to
outperform identical applications  using explicit  messaging systems  on the
same hardware platform.


                                                             Kim Skak Larsen