A Coordinated Location Policy for Load Sharing in Hypercube-Connected Multicomputers

• Authors: Yi-Chieh Chang;Kang G. Shin
• Venue: IEEE Transactions on Computers
• Year: 1995

Uneven task arrivals in a hypercube-connected multicomputer may temporarily overload some nodes while leaving others underloaded. This problem can be solved or alleviated by load sharing (LS); that is, some of the tasks arriving at overloaded nodes, called overflow tasks, are transferred to underloaded nodes. One important issue in LS is to locate underloaded nodes to which the overflow tasks can be transferred. This is termed the location policy. Any efficient location policy should distribute the overflow tasks to the entire system instead of 驴dumping驴 them on a few underloaded nodes. To reduce the overhead for collecting state information and transferring tasks, each node is required to maintain the state information of only those nodes in its proximity, called a buddy set. Several location policies驴random probing, random selection, preferred lists, and bidding algorithm驴are analyzed and compared for hypercube-connected multicomputer systems. Under the random-selection and preferred-list policies, an overloaded node can select, without probing other nodes, an underloaded node within its buddy set, while under the random probing policy and the bidding algorithm the overloaded node needs to probe other nodes before transferring the overflow task. Task collision(s) is said to occur if two or more overflow tasks are transferred (almost) simultaneously to the same underloaded node. The performances of these location policies are analyzed and compared in terms of the average number of task collisions. Our analysis shows that use of preferred lists allows the overflow tasks to be shared more evenly throughout the entire hypercube than the other two location policies.