IOMeans: Classifying Multi-concurrent I/O Threads Using Spatio-Tempo Mapping

Abstract

There is a trend to apply machine learning techniques to improve the performance of core storage components. Typically, a storage system is shared by many concurrent workloads. Because of this, it’s challenging for contemporary machine learning algorithms to learn the hidden patterns from the entangled traces. We develop a novel temporal-aware sequence classification to mine the correlation between I/O requests and represents the addresses with multidimensional vectors that shows better spatial locality. We can efficiently split and clean the entangled I/O trace. By integrating with Recurrent Neural Network (RNN), we greatly improve the cache hit ratio for several concurrent file access workloads.