Merits of a probabilistic approach to properties in process group systems

Abstract

Distributed computing systems have traditionally been designed using either a purely asynchronous methodology, in which no use is made of time, or a purely synchronous one, in which all delays are bounded and clocks are assumed to be synchronized. We argue in favor of an intermediate approach, in which delay states would be systematically removed from an asynchronous system, resulting in a high probability of predictably good performance. Preliminary and very informal analysis suggests that this would yield systems with predictably high performance, and yet that are as likely to achieve this performance as would a real-time system based upon a traditional synchronous design. Such thinking leads to a paradox. On the one hand, we conclude that asynchronous systems may be able to offer real-time properties with high probability. On the other, it fosters concern that all properties of distributed systems might better be viewed as probabilistic.

The author was supported in part by ARPA and ONR under contract number N00014-92-J-1866.