The system data structure contention problem and a novel software solution for shared memory , floating control parallel systems

Of the many varieties of multiprocessor architecture which have been proposed in the last five to eight years, shared memory, floating control multiprocessor systems are in many ways the most elegant. Shared memory, floating control architectures are distinguished by the structural attribute that each processor is an equal partner in the management of system resources even to the point of sharing a common copy of the operating system code and its data structures. A central feature and issue of such systems is the provision of dynamic processor load balancing. The most natural technique, however, for assuring dynamic processor load balancing by dispatching processes from a single shared queue is, we show, highly inefficient. We make the argument, validated by simulation studies, that the contention for the process dispatching queue, as they are commonly implemented, becomes so great so quickly as to severely limit the size and utility of such architectures. Examining the requirements for a solution to this problem, we derive anew, highly concurrent process dispatching queue structure with constant time average performance for all operations. * This research was supported by grant K-5-27116 from United Telecommunications, Inc., Westwood, Kansas.