Co-allocation strategies for fragmented data in grids

Abstract

Data replication is used in grids to increase data availability. A co-allocation strategy in data grids uses parallel data transfer from multiple servers to improve the transfer completion time. The performance of a co-allocation strategy is good if every available server can fully utilize the connection between the server and the client to transfer data. However, because of storage limitation and data ownership, it may not be possible to replicate the whole dataset. As a result, each server on grids can choose to replicate only some portion of a dataset. The replica of a fragment of a dataset is called a fragmented replica. Co-allocation algorithm for complete replica can be modified to allow downloading fragmented replicas. For example, a one-by-one co-allocation algorithm is modified to allow fragmented replica. In this thesis, the dynamic co-allocation algorithm is modified to support fragmented replica, using Random-with-weighted-probability, Biggest-remaining-first and Fewest-replicas-first algorithms. A simulation of a grid is implemented to measure the transfer completion time. From the experiments, Biggest-remaining-first and Fewest-replica-first algorithms perform well when fragments are moderately replicated, i.e., fragments are replicated on half of the servers. Furthermore, an advantage of these algorithms is that they are not sensitive to the variance of the transmission rate.