Special Issue on Multiprocessor Operating Systems 


			  Steve J. Chapin  

		  Department of Computer Science  
		      University of Virginia  
  		    Charlottesville, VA  22903 
 
 

With the advent of high-bandwidth, low-latency communications
networks, the line separating distributed systems and parallel systems
has blurred.  The implementation of message-passing libraries and
software distributed shared memory has muddied the distinction
further.  Still, differences remain between the two camps, and the
three articles in this issue related to our special topic reflect
this.

In our first paper, by Maccabe, Riesen, and van Dresser, describes
features of the Puma operating system.  Puma is an operating system
for massively parallel machines, and is being developed at at Sandia
National Laboratories.  Puma runs on Sandia's 1800-node Intel Paragon
supercomputer and supports grand challenge computational problems such
as large-scale simulations.  In this environment, speed wins, and
Puma's design embodies that credo with highly optimized inter-process
communication.  This paper describes Puma's use of processor modes to
support efficient communication.

Our second paper, by Lumpp et al. at Kentucky, focuses on application
adaptability in large-scale distributed systems.  Their work focuses
on    reactive objects , which allow an application to monitor the
status of the system and to change their behavior to improve
performance.  These behavioral changes may include changes in the
algorithm used to compute problem solutions, or changes in the data
structures used to represent computational objects.  This work is
part of the UNIFY distributed operating system project at the
University of Kentucky.

Our last paper on the special issue topic is an extended abstract by
Fahmy and Heddaya.  Their work is in the context of the    bulk
synchronous parallel  (BSP) programming paradigm, which divides a
computation into a sequence of coarse-grained blocks, called
supersteps, followed by barrier synchronization.  This paper describes
BSPk, a programming model that incorporates BSP and adds two new
features:    communicable memory  and    lazy barriers .  In
addition, their programming model supports both message passing and
distributed shared memory environments.  The result of these features
is hoped to be an improvement in efficiency for implementing programs
using the BSP paradigm.

I would like to thank the TCOS Bulletin editor, Sumi Helal, for giving
me this opportunity to act as a guest editor and for his inexhaustible
patience with missed deadlines.