An Improved Adaptive Neuro-Fuzzy Inference System as Cache Memory Replacement Policy

This paper proposes an improved version of the Adaptive Neuro-Fuzzy Inference System (ANFIS) replacement policy called improved ANFIS Replacement Policy (iANFIS-RP). Previous ANFIS replacement policy shows improvement over least recently used (LRU) at level 2 cache memory but performed worse than LRU at level 1 (L1) in terms of miss ratio. The proposed iANFIS-RP is an attempt to improve the previous ANFIS performance at L1. iANFIS-RP uses recency and frequency of reference to make replacement decision when a miss occurs in a saturated cache memory of microprocessor. The iANFIS-RP has been trained using a new set of data on MATLAB with hybrid learning algorithm. Three triangle MFs for recency and frequency inputs are used. The iANFIS-RP Output values in the set are updated when a miss occurs instead of a hit. Experiment simulations are performed in L1 cache size of 4 to 256 kB with fixed block size and associative using Sim-outorder and eight Standard Performance Evaluation Corporation (SPEC) Central Processing Unit (CPU) 2006 benchmarks. The experiment results show that the iANFIS-RP performed better than the ANFIS replacement policy at various L1 cache sizes. This shows that the modifications have successful improved the previous ANFIS replacement policy.