PERFORMANCE OPTIMIZATION OF PARALLEL PROCESSING COMPUTER SYSTEMS
Abstract
Extensive research has been conducted over the last two decades in developing parallel processing computer systems to achieve high system throughput, availability, reliability, flexibility, and fault tolerance. Array processors that execute a single instruction stream over multiple data streams are extended in this thesis to the emerging field of multiple vector processing and distributed computation. This thesis investigates performance optimization of two classes of parallel processing computer systems. One class is the shared-resource Multiple-SIMD (MSIMD) array processors, and the other is the distributed Multiple Processor System (MPS). In an MSIMD array processor, the optimal size of the resource pool of Processing Elements (PEs) and the sufficient buffer size are systematically determined in this study. A probabilistic optimal scheduling policy is developed to achieve load balancing and minimal average job turnaround time in an MPS. Queueing networks are used in modeling the above two types of parallel processing computer systems. Analytical relationships are established between system parameters and a chosen performance criterion. These analytical results are used to optimize the system performance of parallel processing computers. A queueing model is proposed for MSIMD computer systems. Different solution techniques are presented and compared. Analytical methods to choose the optimal size of the shared-resource pool and to determine the sufficient buffer size are developed. Research results on MSIMD computers can be applied to the design and evaluation of computer systems, such as Illiac IV (the original design), MAP, PM('4), and SCR systems proposed by many computer architects. Optimal probabilistic job scheduling policies are developed for both homogeneous and heterogeneous multiple processor systems. Environments of both single job class and multiple job classes are considered. The probabilistic scheduling algorithm is shown applicable to loosely coupled multiple processor systems, like the MACE system at Purdue Computing Center and the commercial TANDEM/16 computer, in order to achieve load balancing among multiple processors in the context of overall system performance. This algorithm can be also applied to computer communication networks for optimal probabilistic message routing.
Degree
Ph.D.
Subject Area
Computer science
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.