Performance evaluation of the FPS Model 500 Vector Processor

In this paper we present and discuss performance obtained on a FPS 511EA for some vector and matrix operations and for two applicative programs. The FPS 500EA is a 64-bit multiprocessor system with Unix Operating System which can be configured with scalar RISC and vector processors; up to four processors can be installed with a maximum of two vector processors. The vector processor, with 30 nsec clock, an adder and a multiplier, provides a peak performance of 67 MFLOPS. The system available at our site, on which we obtained the presented results, is a FPS 511 EA with one scalar and one vector processor. In the first part of the paper we evaluate the efficiency of the vector architecture and available software tools for some basic simple and compound vector operations and for matrix multiply. When possible different approaches are compared; influence on performance of vector length, stride and memory interleaving is shown and some techniques for achieving a better overlap of memory and arithmetic operations are presented. In the second part we describe some experiences of porting codes from FPS M64/60 to FPS 511EA and compare performance obtained on the two machines. The codes considered are an implementation with matrix kernels of the Back-Propagation algorithm for neural network training and two versions of the Grassberger & Procaccia algorithm for computing the Correlation Dimension, originally developed for maximum performance on FPS M64/60.