[comp.parallel] Self-Synchronizing Concurrent Computing Systems

vijaykm%janus.Berkeley.EDU@ucbvax.Berkeley.EDU (Vijay Madisetti) (10/20/89)

I thank the significant amount of interest in some ERL preprints
on performance analysis of distribuued synchronization in message
passing systems.  I was unable to reply to a large number of
requests.  However, the dissertation that forms the (cleaned
up and better explained) core of the work, is now available
as an ERL Memorandum M89/122 from ERL at UC Berkeley, EECS.
 
                        Ph.D Dissertation
          "Self-Synchronizing Concurrent Computing Systems"

                  Vijay Krishna Madisetti
 Department of Electrical Engineering and Computer Sciences
                  University of California
                     Berkeley, CA 94720
        
                ERL Memo M89/122

                    Abstract

     While the past few years  have  witnessed  an  unpre-
cedented  advance  in  the  status  of computing hardware,
software has not caught up with this pace of  development.
There  has  also been a growing need within the scientific
community for increased access to high performance comput-
ing  engines  for the efficient solution of a wide variety
of compute-bound problems. Our effort has been focused  on
the  development  of efficient algorithms and software for
high-speed parallel scientific computing in an  effort  to
meet this demand.


     This thesis presents theory and design of a new  dis-
tributed  computing  system, the Self-Synchronizing Concu-
rrent Computing Systems (SESYCCS), for efficient solution 
of  a large class of compute-bound scientific problems. In
this thesis, we propose robust models for distributed com-
putation,  and  analyze  the performance of new algorithms
for synchronization.  Concrete  quantitative  results  are
derived  for  systems  with  two  or  more processors. The
effects of communication delay and finite  memory  on  the
computation are also considered in the analysis.

   This thesis establishes that separating  synchroniza-
tion  from  computation has a number of merits, especially
in boosting the efficiency of implementation and  reducing
memory requirements.  It is also proved that simplicity of
asynchronous communication can be preserved in  a  bounded
memory environment.


Application  of  the  algorithms  to  simulation   of
discrete-event  systems  is described and a new algorithm,
Wolf, is proposed and analyzed.

                      -----------

                                   October 17, 1989

This dissertation is available next week as ERL Memo M89/122 from the
Electronics Research Laboratory (ERL)
Electrical Engineering and Computer Sciences
University of California at Berkeley
Berkeley, CA 94720
(415) 643-8885  (attn: Doris)
-----------------------------------------------------------