Performance of fault tolerant networks of workstations

Functional or dataflow models of computation enable a program´s run-time system to determine which portions of a computation must be repeated when faults occur. Straightforward modifications to the run-time system of Cilk 2.0 (a threaded extension of C) enable a network-of-workstations parallel processing system to tolerate fail-stop faults of the individual processors or the network. It is shown in this paper that the overheads needed to provide this fault tolerance are mainly CPU cycles and memory, with very little additional network load being generated in the absence of faults. This makes it feasible to run long computations successfully on the typical networks of workstations found in large organisations, where ownership, control and distribution of the individual processors may be widely distributed