期刊名称：Advances in Electrical and Computer Engineering
印刷版ISSN：1582-7445
电子版ISSN：1844-7600
出版年度：2014
卷号：14
期号：1
页码：15-24
DOI：10.4316/AECE.2014.01003
出版社：Universitatea "Stefan cel Mare" Suceava
摘要：To date, no cache memory replacement policy that can perform efficiently for all types of workloads is yet available. Replacement policies used in level 1 cache memory may not be suitable in level 2. In this study, we focused on developing an adaptive neuro-fuzzy inference system (ANFIS) as a replacement policy for improving level 2 cache performance in terms of miss ratio. The recency and frequency of referenced blocks were used as input data for ANFIS to make decisions on replacement. MATLAB was employed as a training tool to obtain the trained ANFIS model. The trained ANFIS model was implemented on SimpleScalar. Simulations on SimpleScalar showed that the miss ratio improved by as high as 99.95419% and 99.95419% for instruction level 2 cache, and up to 98.04699% and 98.03467% for data level 2 cache compared with least recently used and least frequently used, respectively.
