Credit scheduling and prefetching in hypervisors using Hidden Markov Models

The advances in data storage technologies like Storage Area Networking (SAN), virtualization of servers and storage, cloud computing have revolutionized the way the data is stored. A large number of business organizations, universities, hospitals, research organizations are now deploying SAN, not as a luxury but as a necessity. Scientific research organizations like NASA process terabytes of data every day. Accurate analysis and processing of the experimental data call for a need to efficiently store and retrieve the data to and from data storage media. Similarly social websites like YouTube, FaceBook handle large amounts of data every minute. So, the robust performance of any computing and retrieval applications demands a reduction in the latency of data access. Hidden Markov Models (HMM) have been successfully used by researchers to predict data patterns in the areas of speech recognition, gene prediction, cryptanalysis etc. The goal of this research is to reduce the scheduling delay in hypervisors and the latency of reading blocks of data from the disk array using Hidden Markov Models (HMM) in a server virtualized environment. HMM was implemented to identify patterns of read requests issued and exploited to reduce the overall read response time of a server. A Gaussian HMM is used to reduce the scheduling delay and a discrete HMM is used to reduce the read response time. Results observed using HMM were very promising compared to results without HMM in decreasing the overall latency in data access.