Asynchronous Procedures for Parallel Processing

Virtually all the algorithms now being used for network simulation, tracking and optimization are synchronous. They cannot be broken into parallel processes for concurrent execution without inserting a number of synchronization points. The first process to reach a synchronization point must wait for the others to catch up. The resulting delays and communication overheads often grow rapidly with the number of processors. When this happens, large numbers of processors can be used with advantage only if the synchronous algorithm is replaced with an asynchronous procedure. This is not always easy. Asynchronous procedures tend to be counter intuitive, difficult to construct and difficult to analyze. To reduce these disadvantages we will develop concepts useful in thinking about asynchronous procedures, discuss a model after which they may be patterned, introduce some new convergence results and presents some illustrations of how to construct asynchronous procedures. Finally, we will discuss an important application area-contingency constrained optimum flows.