AHPCRC Researchers Demonstrate Heterogeneous Computing at Supercomputing 94

摘要

Recent years have witnessed a vast increase in the number of computing platformswhich are available to computational scientists. A multitude of design paths taken by the computer manufacturers in their never-ending quest for high performance and rapid programmability have resulted in a wide spectrum of CPU, memory and storage architectures. The choice of the optimal computational platform and algorithm for a given problem on the part of the numerical analyst has become much more important than in the past. While many applications significantly benefit from their adaptation to parallel distributed memory computers, some algorithms are inherently non-parallelizeable. Some algorithms, when moved from a vector supercomputer to a parallel architecture which they cannot fully exploit, suffer major degradation in performance. In this case the decision to continue the utilization of such algorithm on a traditional supercomputer may be the correct one. Large, complex, simulations may require many different algorithms to perform the analysis, with sections of the code ranging from embarrassingly parallel to unavoidably sequential. It is here that the concept of heterogeneous computing may provide a welcome alternative to months needed to replace the sequential code, or to leaving the entire package, including the easily parallelizeable portions, on a non-parallel machine. In heterogeneous computation, many kinds of computers may cooperate, with each of the machines performing tasks for which it is particularly suited. The platforms exchange data utilizing a fast network connection.