UNIPROCESSOR PERFORMANCE ENHANCEMENT WITH ADDITIVE SCHWARZ PRECONDITIONERS ON ORIGIN-2000

Most current-day microprocessor designs are characterized by fast proc essors and slow main memory. The performance gap is usually reduced by using small and fast cache memory as a buffer to main memory. In orde r to achieve good performance, it is necessary for an algorithm to red uce cache misses. In this paper, we study the performance characterist ics of several basic iterative kernels on a single microprocessor of t he SGI Gray Origin 2000 system. For large problem sizes, the performan ce of the global preconditioned conjugate gradient (PCG) method is lim ited by frequent access to slow main memory. On the other hand, the ad ditive Schwarz PCG method shows much better performance because of eff ective utilization of the cache memory. (C) 1998 Elsevier Science Limi ted. All rights reserved.