GYRO PERFORMANCE ON A VARIETY OF MPP SYSTEMS

GYRO is a code used for the direct numerical simulation of plasma microturbulence. It is the most physically comprehensive of all such codes, worldwide. It has been ported to a variety of modern MPP platforms including a number of commodity clusters, IBM SPs and the Cray X1. The performance and scaling of GYRO on many of these systems has been previously documented, and exhaustive testing and performance analysis has already been the subject of numerous presentations and publications. Although we briefly describe the mathematical structure of the equations solved by GYRO, we prefer to focus on the data layout and redistribution algorithms. Performance comparisons of different platforms are detailed, with emphasis on a section of code which requires very large communication bandwidth. In addition, we discuss the performance of a recently implemented FFT algorithm for the evaluation of the nonlinear terms. This algorithm gives a performance improvement for very large grid sizes. However, because the FFT vector length very rarely exceeds 192 in practice, one never operates the code in a regime where the full aymptotic dominance of the FFT over the original direct method is realized.