Research On Replica Management Strategies In Data Grids

Replica Management is one of the critical parts in Data Grids. Replication can reduce access latency, improve data locality, and increase robustness, scalability and performance, or balance the loads for distributed applications. Excellent replica management strategies are important to improve the Qos in data grids. This paper investigates on four compenent of replica management in data grids: replicas creation strategies, replicas selection strategies, replica location mechanism and scheduling the replicas requests on the home node.(1) Replicas creation strategies tend to make decisions on where and when to create which data replicas. In this paper, we propose our replicas creating strategy based on investigating and analyzing the related researches. The method integrates the clients-driven dynamic Cascading replication model and the profit-driven strategy. It is prudential when deciding to create replicas on one grid node while considering two factors: the client and the profit. The strategy pays more attention to the future value of the data replica.(2) Replicas selection strategies discuss how to select the best replicas among the current known replicas. After investigating and analyzing the related works on the replicas selection, we bring forward the strategy based on open auction protocols in the economy marketplace employing the token technique. We also evaluate the method in the article: Utilizing the token technique, the method based open auction protocols with the properties of quickness, briefness and efficiency.(3) Replica location is also important to reduce access latency in data grids. In this paper, we proposed the replica location mechanism based on DHT and the Small World theory in process of analyzing the advantages and the disadvantages of the related researches. Forming Virtual Organizations based on the Small World theory by the VODL (Virtual Organization Description Language). By checking the VODL, the grid node joins into a VO based on the mutual principle. After constructing the model by the Vos, we employ the different methods between the inner domain and the inter domain. In the inner domain, Gossip technique is utilized to spread the replicas information within the domain. It is the VOServers' responsibility to locate the physical data element. While among the inter domains, DHT technology is employed to spread the replicas information on the home nodes, who will see after the replicas location. Simulation experiment results show that loads on the home nodes are balanced when employing the DHT technique. The replica location mechanism-Ridrop (drop the replica information) with the properties of dynamic self-adaptability as well as reliability, scalability and decentralization.(4) Queuing phenomenon occurs on the home nodes when there are many grid jobs request to locate data replicas. Employing efficient algorithms to schedule replica location requestes to the home nodesthe can improve the performance of the system. According to the characteristics of heterogeneous of the data grids and the different process capability of the home node, and borrowing the idea of the general M/M/N queuing theory, this paper proposed the scheduling algorithm based on FCFS and "able person should do more work". The experimentsshow that when employing the scheduling algorithm, the simulation system perform better than employing the scheduling algorithm based on FCFS in the M/M/N queuuig model on the assumption that the speed of generating the replica location request or the total ability of the home nodes is equal between the two scheduling algorithms. Meanwhile, the wider discrepancy on the location capability of the home nodes, the scheduling algorithm based on FCFS and "able person should do more work" can achieves the better performance for the simulated systems, which is very important to rapidly locate the replicas in data grids.In conclusion, after investigating, on the related research on the replica management, we propose our own methods or strategies on replicas creation, replicas selection, replica location mechanism and scheduling algorithms on the home nodes. We evaluate and analyze advantage of the replicas creation and selection strategies. When come to the latter two components, simulation experiments done on them show that our mechanism or algorithm achieves better performance than the others related.