 |
|
| |
|
The gap between sustained and peak performance of scientific applications is a well-known problem in high performance computing. The recent development of parallel vector systems offers to bridge this gap for a significant fraction of computational science codes and deliver a dramatic increase of computing capabilities. This phenomenon was dramatically highlighted when the Japanese Earth Simulator System was recently announced. The Earth Simulator project itself was not news. What was surprising, was that 87% of the theoretical peak speed was achieved at such an early stage in its development. More important than the raw performance though is what this new capability entails for the climate modeling community and other scientific communities that rely on modeling and simulation.
However, effectively programming these modern parallel vector architectures will be more difficult than on previous parallel vector machines, due to the multiple levels of parallelism that must be exploited in order to achieve high performance. The goal of this study is to answer strategic questions regarding the expected performance of DOE high-end computing applications on the latest generation of vector multiprocessors: the NEC SX6, the Cray X1, and the Japanese Earth Simulator System (ESS). We will study key factors of these modern parallel vector systems, including runtime, scalability, programmability, portability, and memory overhead while identifying potential bottlenecks. The results gathered in this study will benefit the scientific computing community in general and help prepare DOE and NERSC for future system acquisitions.
|
