Buffer cache management: predicting the future from the past

Abstract

Efficient and effective management of the buffer cache in the operating system becomes increasingly important as the speed gap between microprocessors and hard disks becomes wider This paper presents different techniques for predicting the future disk access patterns from the access history of each block and the access patterns detected for related blocks. The first part of the paper focuses on a block replacement policy called LRFU (least recently/frequently used) that subsumes the well-known LRU (least recently used) and the LFU (least frequently used) policies. Then, the next part discusses techniques for handling regular references such as sequential and looping references. Finally, the results from both trace-driven simulations and our implementation of the techniques within a real operating system are presented.