Date of Award
Doctor of Philosophy in Computing Sciences - (Ph.D.)
Computer and Information Science
James A. McHugh
Alexandros V. Gerbessiotis
The development of computers with hundreds or thousands of processors and capability for very high performance is absolutely essential for many computation problems, such as weather modeling, fluid dynamics, and aerodynamics. Several interconnection networks have been proposed for parallel computers. Nevertheless, the majority of them are plagued by rather poor topological properties that result in large memory latencies for DSM (Distributed Shared-Memory) computers. On the other hand, scalable networks with very good topological properties are often impossible to build because of their prohibitively high VLSI (e.g., wiring) complexity. Such a network is the generalized hypercube (GH). The GH supports full-connectivity of its nodes in each dimension and is characterized by outstanding topological properties. In addition, low-dimensional GHs have very large bisection widths. We propose in this dissertation a new class of processor interconnections, namely HOWs (Highly Overlapping Windows), that are more generic than the GH, are highly scalable, and have comparable performance. We analyze the communications capabilities of 2-D HOW systems and demonstrate that in practical cases HOW systems perform much better than binary hypercubes for important communications patterns. These properties are in addition to the good scalability and low hardware complexity of HOW systems. We present algorithms for one-to-one, one-to-all broadcasting, all-to-all broadcasting, one-to-all personalized, and all-to-all personalized communications on HOW systems. These algorithms are developed and evaluated for several communication models. In addition, we develop techniques for the efficient embedding of popular topologies, such as the ring, the torus, and the hypercube, into 1-D and 2-D HOW systems. The objective is to show that 2-D HOW systems are not only scalable and easy to implement, but they also result in good embedding of several classical topologies.
Wang, Qian, "A new-generation class of parallel architectures and their performance evaluation" (1999). Dissertations. 992.