Reliable designs for massively connected multiprocessors
Reliable designs of some massively connected multiprocessors are investigated from the following aspects: (1) For SIMD machines with N=2{sup n} pairs of PEs (processing element) and MMs (Memory Modules) interconnected by a multi-stage interconnection network or crossbar, a simple and efficient approach for improving the reliability has been proposed. The approach combining fault detection, fault tolerance and highly accurate computation requires very little extra cost for a reliable design. By running a particular dual version, the whole system can be viewed as two duplicates with half of the PEs in each so that it can continue to operate even with up to n faulty PEs. An efficient self-diagnosis method for detecting all faulty PEs in time O(log{sup 2}N) as well as a highly reliable computation configuration is presented. (2) For Hypercube-like MIMD machines, reliable design is a more complicated issue. The author has first proposed an enhanced hypercube called multi-virtual hypercube (MVH) which can be easily built based on today's VLSI technologies used for hypercubes. Its topological properties have been investigated. It has been compared with the Hypercube and has been shown to have a much shorter diameter, smaller average delay, more independent paths and more subcubes. Based on the MVH, a new class of topologies called Substituted and Modified Hypercube (SMH) has been proposed. An SMH with degrees ranging from 3 to log{sub 2}N has many advantages such as shorter diameter, improved robustness and easy routing. Some algorithms have been investigated. A new parallel matrix multiplication algorithm has been proposed for Hypercube systems. With this algorithm the hypercube or the newly proposed systems can tolerate more faulty PEs.
- Research Organization:
- Minnesota Univ., Minneapolis, MN (USA)
- OSTI ID:
- 6090157
- Country of Publication:
- United States
- Language:
- English
Similar Records
Embeddings, communication and performance of algorithms in faulty hypercubes
Processor allocation and communication in networks
Related Subjects
990200* -- Mathematics & Computers
ALGORITHMS
ARRAY PROCESSORS
COMPARATIVE EVALUATIONS
COMPUTERS
DESIGN
DIGITAL COMPUTERS
ELECTRONIC CIRCUITS
FAULT TOLERANT COMPUTERS
HYPERCUBE COMPUTERS
INTEGRATED CIRCUITS
MATHEMATICAL LOGIC
MATHEMATICS
MICROELECTRONIC CIRCUITS
RELIABILITY
TOPOLOGY