Cache performance and algorithm optimization

A technique to enhance the cache performance of some blocked algorithms is proposed. According to the results of number theory, the author presents a principle for array padding so that accesses of array subblocks do not generate conflict misses. The technique is used to calculate LU factorization and matrix multiplication. The principle is tested on a shared memory multiprocessor. The practical results agree with the theoretical analysis, and 20% to 150% increasing in performance is achieved.