Cache Replacement Based on LSTM in the Second Cache in Virtualized Environment

Abstract

Virtualized environments have become one of the most important platforms in current information systems. In the cases of a virtualized environment using both guest and host operating systems, two caches in the guest and host operating systems work concurrently and improvement of hit ratios of both caches are important. However, the second cache, which is the host operating system cache, has a negative locality of reference and the most popular cache replacement algorithm called LRU (least recently used) does not work well. For this issue, we propose to apply LSTM (Long short-term memory), which is a deep learning network for time series data, for cache replacement in the second cache. First, we explain a famous issue in the second cache. Second, we propose an LSTM-based cache replacement. Third, we evaluate this method with our prototype implementation and show that our method can improve I/O performance.