Autonomic Storage System Based on Automatic Learning

Abstract

Abstract. In this paper, we present a system capable of improving the I/O performance in an automatic way. This system is able to learn the behavior of the applications running on top and find the best data placement in the disk in order to improve the I/O performance. This system is built by three independent modules. The first one is able to learn the behavior of a workload in order to be able to reproduce its behavior later on, without a new execution. The second module is a drive modeler that is able to learn how a storage drive works taking it as a "black box". Finally, the third module generates a set of placement alternatives and uses the afore mentioned models to predict the performance each alternative will achieve. We tested the system with five benchmarks and the system was able to find better alternatives in most cases and improve the performance significantly (up to 225%). Most important, the performance predicted where always very accurate (less that 10% error).

A Systematic Scheme to Resolve QoS Dissatisfaction for Storage Cluster

“…Previous research mainly stressed the problem of scheduling each migration from its original location to its new one to minimize the total migration time [7]. Little research exist to create an efficient data migration plan, for example, to maximize the number of clients that can be served by the parallel disks or to automatically improve storage I/O performance [8].…”