Efficient Pair of Replacement Algorithms for L1 and L2 Cache for Matrix Multiplication

Abstract

Processors speed is much faster than memory; to bridge this gap cache memory is used. This paper proposes a preeminent pair of replacement algorithms for Level 1 cache (L1) and Level 2 cache (L2) respectively for the matrix multiplication (MM) application. The access patterns of L1 and L2 are different, when CPU not gets the desired data in L1 then it goes to L2. Thus the replacement algorithm which works efficiently for L1 may not be efficient for L2. With the reference string of MM, the paper has analyzed the behavior of various existing replacement algorithms at L1 and L2 respectively. The replacement algorithms which are taken into consideration are: least recently used (LRU), least frequently used (LFU) and first in first out (FIFO). This paper has also proposed new replacement algorithms for L1 (NEW ALGO1) and for L2 (NEW ALGO2) respectively for the same application. Analysis shows that by applying these algorithms at L1 and L2 respectively miss rates are considerably reduced.