Load sharing in hypercube multicomputers in the presence of node failures

Two important issues associated with load sharing (LS) in hypercube multicomputers are discussed and analysed: (i) ordering fault-free nodes as preferred receivers of overflow tasks and (ii) developing an LS mechanism to handle node failures. The authors previously (1989) proposed to order the nodes in each node´s proximity into its preferred list of receivers for the purpose of LS in distributed real-time systems. However, the occurrence of node failures will destroy the original structure of a preferred list if the failed nodes are simply dropped from the list. Three algorithms are proposed to modify the preferred list to retain its original features. Based on the modified preferred lists, node failures can be tolerated by equipping each node with a backup queue which stores and updates the arriving/completing tasks at its most preferred node. Simulation results show that this approach, despite its simplicity, can greatly reduce the number of task losses compared to approaches that do not use backup queues.<>