Parallel Matrix Computations.
Abstract
The purpose of this effort is to develop realistic algorithms for matrix computations on parallel computers. It has been long observed that the usual algorithms of numerical linear algebra contain a great deal of inherent parallelism. For example, if the arithmetic operations that can be performed in parallel in Gaussian elimination are actually so executed, the time to decompose an nXn matrix is reduced from cu n to n only recently, with the emergence of cheap, small microcomputers, has it become feasible to exploit this parallelism on anything but a trivial scale there is under development a parallel system, called ZMOB, consisting of 256 micro-processors connected on a conveyor belt. This belt is so fast and its architecture is such that any two processors can communicate without interfering with the communications of other pairs of processors. Thus the ZMOB is an ideal tool for simulating an arbitrarily connected network of computers. This feature of the ZMOB is particularly useful in investigating parallel matrix algorithms. As was noted above, there is much parallelism in most current matrix algorithms. However, to exploit it, information must be moved from processor to processor. This constitutes the chief bottleneck in parallel matrix algorithms; interconnections between processors are expensive, and in a practical system one can assume only a limited amount of connectivity. The ZMOB provides a means of testing and comparing different types of interconnections, since all one has to do is use the rich connections provided by the ZMOB conveyor belt.
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 11, 1985
- Accession Number
- ADA159252
Entities
People
- D. P. O'leary
- Gilbert W. Stewart
Organizations
- University of Maryland