Design methodology and heuristic for reconfigurable hybrid macro pipeline multiprocessor

Abstract

"The research developed a reconfigurable hybridized multiprocessor system, Hybridized Macro Pipeline Multiprocessor (HMPM), which combines two multiprocessing techniques –macro pipeline and parallelism. This solves the problems associated with using multiprocessor system to execute sequential applications. In addition, a heuristic is developed for the system, this reconfigures the processors in each cluster according to input load in order to prevent load imbalance. Multiprocessor system has led to increased throughput of system in many applications such as image processing, computer graphic and real time systems. It has also contributed to the increasing use of parallel hardware along-with the associated software, which solves the bottleneck associates with high computation tasks. There are different ways of increasing the throughput in a multiprocessing environment. Parallelism technique; where different processors are allocated different parallel-subtasks at the same time. In macro pipeline, all the available processors must sequentially work on all the subtasks. Several researchers have proposed different form of multiprocessors using any of these multiprocessing techniques. However, these techniques are application dependent. That is some applications are sequential, for these applications, parallelizing their execution will reduce the throughput. Some are parallel in nature, executing these sequentially will also reduce the throughput. This made some of the proposed multiprocessor system application-dependent. This paper proposes a reconfigurable Hybridized Macro Pipeline Multiprocessor (HMPM); a hybrid multiprocessor system that combines sequential and parallel execution techniques in order to increase the multiprocessor throughputs and make it application-independent. HMPM uses homogenous processors as the processing elements. A streaming application is taken and broken into a series of subtasks, which are parallel assigned to different pipelined clusters. The problem of mapping each of the streaming unit into each of pipeline stage is solved with a simple architecture and heuristic that efficiently determine the design space.

"