Performance Evaluation of Data Migration Policies for a Distributed Storage System with Dynamic Tiering

Abstract

Recent years, even client nodes can easily have enough storage area owing to the spread of inexpensive storage devices. In many cases, the large amount of an unused storage area is left on storage devices over the network. It is important to utilize such fragmented unused storage area. Hence, we have for our object to build an environment which not only file servers but also client nodes can provide their surplus storage area as public storage. However, such computing environment has essentially heterogeneity. The heterogeneity of the system causes difficulty for properly allocating data blocks onto the storage devices. We consider that system heterogeneity is brought by even a dynamic aspect. It means that both a static aspect (e.g. differences of hardware configuration) and a dynamic aspect (e.g. imbalance of I/O workload) make such heterogeneity. This dynamic heterogeneity is not able to be solved statically. Therefore, data blocks should be located on a suitable storage device according to their access patterns at runtime. For example, frequently accessed data blocks are better to locate on high-performance storage, and it may cause no inconvenience when rarely accessed data blocks are located on low-performance storage. However, such location optimization in accordance with access patterns of data blocks makes the burden too heavy for system administrators.