As part of the HORIZON project currently underway at the Supercomputing Research Center, a set of application programs are being written and their performance is being evaluated. This paper discusses two of these applications: the fast Fourier transform and sparse matrix computations. For both problems, we develop efficient implementations that take advantage of HORIZON'S many unique features, including fined-grained synchronization, multiple instruction streams, and a horizontal instruction set. Our results indicate that HORIZON will offer very high performance, sometimes in excess of a sustained rate of one floating point operation per tick.
IntroductionThe HORIZON supercomputer is a 256 processor shared memory MIMD machine that is currently being designed and evaluated at the Supercomputing Research Center. As part of the evaluation of HORIZON, a set of application programs are being written and simulated to determine their potential performance. These application programs have been chosen to represent a variety of computationally intensive tasks that normally demand supercomputer performance.In this paper, we give a flavor of these application studies by discussing work that has been done to date on the fast Fourier transform and sparse matrix computations. Our results indicate that HORIZON will indeed be a high performance machine capable of outperforming existing supercomputers by at least two, possibly even three orders of magnitude.For both problems, we show how to map efficient algorithms effectively onto HORIZON in order to obtain high performance. A key part of this mapping process is the creation of multiple Istreams for each HORIZON processor, which allows the delay in accessing memory to be hidden in the algorithms. We also discuss how to take advantage of some of the novel aspects of the HOR-IZON instruction set in the code for our algorithms.For the fast Fourier transform, we have implemented an algorithm based on the radix4 Stockham variant that promises excellent performance. At each stage of the algorithm, multiple Istreams perform independent radix4 computations. Each I-stream has enough registers to support a radix-4 computation without spilling to memory in the middle, which increases performance
