Distributed System Behavior

We propose a new view of a distributed system behavior that can be used to better understand a distributed computation, better specify the properties of various services implemented in these systems, and identify conditions under which some of the important problems such as consensus are solvable. Most existing distributed systems are not completely synchronous, i.e. there are periods when communication delays experienced by messages exchanged may be longer than the time out delay associated with them. As a result, some of the important functionalities such as consensus and real-time responsiveness cannot be guaranteed in such systems. This paper makes an attempt to better understand such systems by observing their behavior as a computation progresses. A distributed system behavior is viewed as alternating between periods of asynchronous and synchronous behaviors. The synchronous and asynchronous behaviors are defined over a real-time interval in terms of the success or failure of communication between processes within a known time bound. This view of distributed system behavior is more realistic than that has been done in the past, where a system is assumed to be either completely synchronous or completely asynchronous. We believe that a majority of the existing distributed systems behave synchronously most of the time with short periods of asynchrony in between.

Publications

S. Mishra, Distributed System Behavior in the Presence of Asynchrony and Failures. Proceedings of the 11th International Conference on Systems Engineering, Las Vegas, NV (July 1996) 420--425.
S. Mishra, An Experimental Study of Distributed System Behavior. Proceedings of the 12th ISCA International Conference on Computers and their Applications, Phoenix, AZ (March 1997).