Minimization of supervisor conflict for multiprocessor computer systems

• Authors: Randy J. Raynor;John M. Gwynn, Jr.
• Venue: ANSS '76 Proceedings of the 4th symposium on Simulation of computer systems
• Year: 1976

In a multiprocessor system, the handling of interrupts generated by jobs in the processors is assigned to a supervisory program and associated data base. The two basic philosophies for deciding which processor executes the supervisor are master-slave and floating executive control. In either case, queueing of requests to the supervisor may occur. With the master-slave structure, the master processor can handle only one request at a time. With floating executive control, while any processor can execute the supervisor, only one processor at a time can be allowed to access the supervisor's data base. A processor which is waiting to use the supervisor is not doing useful work. Therefore, queueing of requests to the supervisor causes a degradation in the performance of the system. Previous studies have indicated that this degradation is significant for systems with a large number of processors. A methodology is developed through simulation which reduces supervisor queueing by scheduling jobs to processors such that they request the use of the supervisor when the supervisor is predicted to be available. Information concerning job execution characteristics required by the scheduler is shown to be available at sufficient accuracy via dynamic monitoring and forecasting during normal program execution.