Runtime Data Declustering based on Bandwidth-on-Demand and its Evaluation over SAN-connected PC cluster

Clusters of computers are used in large scale server sites recently, because of their good scalability and cost/performance ratio. In addition, Storage Area Network (SAN) is introduced in order to consolidate back end of such systems. I/O-bottleneck is serious problem in such an environment, because some important data-intensive applications often access part of data concurrently and repeatedly through SAN. In order to resolve this problem, an idea of “Bandwidth-on-Demand” is introduced in this paper. Runtime declustering middleware based on this idea is implemented on SAN-connected PC cluster, and evaluated with data mining application. According to the results of experiments, an overhead of this middleware is trivial, and I/O-bottleneck problem is resolved after declustering is completed, therefore, preferably good performance improvement is achieved. The runtime declustering middleware achieves better performance improvement especially when the number of nodes used in the cluster is large, because I/O-bottleneck problem becomes serious in such a case. Furthermore, data migration after declustering costs minimal overhead.