The Prediction of Forward Reuse Distance Using a Classification Model for Write Buffer Management
Hyejin Cha, Taeseok Kim · Asia-pacific Journal of Convergent Research Interchange · 2023
The SSD write buffer is used to improve the latency of I/O requests and the lifetime of flash memory.Most algorithms for managing the SSD write buffer use past reference information such as LRU (Least Recently Used) or predict the future reference using deep learning with high computation cost.In this paper, we propose a technique for managing write buffers by efficiently predicting future reference patterns with a reasonable computation overhead using machine learning.If the future reuse distance, which means when the same data is re-referenced, is greater than the write buffer size, a strategy of directly flushing it in the flash memory without storing it in the buffer is used.However, since it is impossible to accurately predict the future reuse distance and is not essential for buffer management, we first group the future reuse distance into several levels considering the size of the write buffer.Then, a priority-based buffer management policy is proposed by taking the grouped future reuse distance as a priority for caching.As a result of the simulation, the proposed method showed a performance close to the optimal Belady's algorithm in terms of hit rate, while showing significant performance improvement compared to the existing LRU-based algorithm.