Petaflops computing: planning ahead

This talk considers the problem of defining success criteria for petaflop computers. Current expectations for teraflop systems show an alarming acceleration of a trend we have seen for many years in high performance computers. Namely, it is becoming increasingly difficult to effectively use the computational capability of these machines. If this situation is not reversed quickly, the term "petaflop computer" may simply mean the next fastest computer that we cannot use. In many cases, we have some understanding of why we cannot achieve anywhere near the peak performance of these machines on real applications. Effective use of these resources is a highly complex optimization problem that must be solved over all of the different components of each application program. Given this complexity, it is the responsibility of our community to better quantify our progress in developing high perforrnance systems with more meaningful metrics than simply "peak floating point operations per second." We need to develop metrics and tools that help us to enhance the end-to-end performance of solving large scientific applications on these advanced machines