Intf-HybridMem: uma abordagem para gerência de memórias híbridas explorando sistemas Fuzzy

Abstract

In recent years, the area of computer architecture has watched the consolidation of non-volatile memory technologies (NVMs) as a promising alternative for employment in computer systems. NVMs have behavioral characteristics that differentiate them from DRAM, having superior performance in the execution of read instructions, but lower during the execution of write instructions. In addition, the average endurance of NVMs tends to be lower than DRAM, since write operations, during use, introduce a potential degradation of its components. Thus, a trend identified for the exploration of NVMs throughout the study and research efforts of this Thesis is its use combined with volatile memories, constituting what is called in the literature of the area as hybrid memories. The central point for the exploration of hybrid memories is the strategy adopted for its management, where it is necessary to decide which data should be stored in which technology. This decision must be based on the data access profile, which must be previously collected and stored. To this end, structures are needed to collect these access profiles. However, these management structures have an inherent size limitation, which leads to the loss of part of the information on access profiles, adding an uncertainty component to the decision-making process. In this scenario, this Thesis has as a general objective the conception of an approach for the management of hybrid memories, called Intf-HybridMem, which explores the treatment of the uncertainty present in the information of the access profiles for decision making. The Intf-HybridMem approach was evaluated considering usage scenarios, designed to include its hardware and software modules. The evaluations indicate the feasibility of using fuzzy systems to recommend migrations in hybrid memories, considering information on the frequency of writes, frequency of reads, and access time. The results obtained confirm the premises of the conceived approach and point to the continuity of the research work.