Dynamic Multimethod Communication between Clusters of Workstations

– As parallel library standards make it easier to use clusters of workstations for large computational jobs, concentration must be placed on the underlying communication mechanism used by the library. Also, with dynamic resource allocation schemes and task migration, it is now possible to move tasks in mid–execution to exploit architectures that offer faster communication networks. This paper describes a distributed execution environment known as Hector that has been constructed to allow for dynamic resource allocation and performance gain resulting from dynamic switching of the communication mechanism. The ability to change the communication mechanism, for example, from Ethernet to shared memory in mid–execution can optimize parallel program task placement to take advantage of the fastest available communication network. The mechanism that permits switching between shared–memory and sockets in mid–execution (as tasks migrate) has been implemented and tested and its performance quantified.