High Performance Computers and Algorithms From Linear Algebra

This chapter discusses high performance computers and algorithms from linear algebra. As multiprocessor designs proliferate, research efforts should focus on “generic” algorithms that can be easily transported across various architectures. If a code has been written in terms of high level synchronization and data management primitives, that are expected to be supported by every member of the model of computation, then these primitives only need to be customized to a particular realization. A very high level of transportability may be achieved through automating the transformation of these primitives. The benefit to software maintenance, particularly for large codes, is in the isolation of synchronization and data management peculiarities. One should be prepared to give up a marginal amount of efficiency in trade for reduced man power requirements to use and maintain software. There are many possibilities for implementation of the general ideas, which are briefly described in the chapter.