Performance Prediction and Calibration for a Class of Multiprocessors
IEEE Transactions on Computers
A multiple processor system for real time control tasks
ISCA '82 Proceedings of the 9th annual symposium on Computer Architecture
IEEE Transactions on Computers
A Cluster Structure as an Interconnection Network for Large Multimicrocomputer Systems
IEEE Transactions on Computers
A Modular Approach to Real-Time Supersystems
IEEE Transactions on Computers
| Hi-index | 0.01 |
An analysis for a multiprocessor system with a shared bus is given. The analysis applies to application areas where the task to be performed can be partitioned into largely independent subtasks. Each subtask requires cyclic execution on continuous data input. The objective is to determine the processing power of such a system as the number of subtasks or, equivalently, as the number of processors is increased. First, it is assumed that the partitioning of the task can be done without incurring any overhead in processor execution time (E) or bus usage time (I). Upper bounds on the processing power are determined that are independent of the bus type and bus allocation scheme. The upper bounds are functions of the number of processors and the ratio (E/I). The upper bound on processing power increases linearly with the number processors until it reaches the value E/I. Secondly, it is assumed that the partitioning causes an overhead in processor execution time and bus usage time that increases linearly with the number of processors. If the processor execution overhead predominates, then the number of processors required to obtain maximum processing power is greatly increased and the maximum processing power attainable is limited by the overhead factor for any E/I. If the bus usage overhead predominates, then the maximum attainable processing power is reduced and is reached with fewer processors, i.e., the bus is saturated with fewer processors. Thirdly, a specific bus allocation scheme is analyzed namely FIFO, assuming randomly distributed processor execution and bus usage times with average values E and I, respectively. In this case, the processing power is lower than what is ideally attainable (partitioning without overhead) for any value of n. As n becomes large, the maximum approached is the same as in the ideal case.