Porting the 3D Gyrokinetic Particle-in-cell Code GTC to the CRAY/NEC SX-6 Vector Architecture: Perspectives and Challenges

Abstract

The impressive performance achieved in 2002 by the Japanese Earth Simulator computer (26.58 Tflops, 64.9% of peak) [1] has revived the interest in vector processors. Having been the flagship of high performance computing for more than two decades, vector computers were gradually replaced, at least in the US, by much cheaper multi-processor super-scalar machines, such as the IBM SP and the SGI Origin series. Although the theoretical peak performance of the superscalar processors rivals their vector counterparts, most codes cannot take advantage of this and can run only at a few percent of peak performance (½¼%). When properly vectorized, the same codes can, however, reach over 30 or even 40% of peak performance on a vector processor. Not all codes can achieve such performance. The purpose of this study is to evaluate the work/reward ratio involved in vectorizing our particle-in-cell code on the latest parallel vector machines, such as the CRAY/NEC SX-6, which is the building block of the very large Earth Simulator system in Japan [2]. This evaluation was carried out on the single node (8 cpus) SX-6 located at the Arctic Region Supercomputing Center (ARSC). Early performance results are compared to the same tests performed on the IBM SP Power 3 and Power 4 machines, on which our particle-in-cell code, GTC, is normally run.