SSD의 성능 및 수명 향상을 위한 데이터 수정 주기 예측 모델 가속화 방안

Abstract

An SSD (Solid State Drive) is a storage device based on NAND Flash Memory. Due to the discrepancy between the write and erase units, garbage collection (GC) must run to delete invalid data. This process duplicates valid data to other blocks, resulting in more writes than requested, which leads to Write Amplification (WA). The Write Amplification Factor (WAF) is the ratio of the actual written amount to the requested write amount, and reducing it is a critical research topic for SSD. Grouping data with similar update periods can reduce WA. Therefore, predicting the update periods of data and organizing them into groups is considered an important research direction. However, the future data update period cannot be known during the processing of wriring request. Recently, several studies have been conducted using machine learning to predict the update periods. Nevertheless, predicting update periods through machine learning can increase latency for SSD due to inference time. This study proposes offloading the computation of update period prediction models to FPGA to reduce inference latency caused by machine learning. Three workloads, including FIO, TPC-C, and Mail, are used in comparing no prediction of update periods, prediction using CPU, and prediction using FPGA. The results show that the inference time with FPGA was 1,244 times faster than with CPU alone. WAF and I/O latency were also measured using an SSD simulator with machine learning applied, demonstrating that reducing WAF through a machine learning model for workload characteristics can enhance SSD performance and lifespan.