Deadlock - Free Routing in a Faulty Hypercube

In massively parallel computers, processors are often connected in a hypercube configuration. Each vertex in the hypercube represents a processor, and each edge represents a communication link. One problem in such a system is avoiding deadlock, a state where there is a cycle of processors, each waiting on the next indefinitely. A second problem is that in a system with many processors, some are bound to fail, leaving a faulty hypercube. This thesis investigates the problem of efficiently moving data between processors in a faulty hypercube while avoiding deadlock. This work was supported by the US Army through grant DAAH04-95-1-0607 and by DARPA under contract N00014-95-1-1246. Thesis Supervisor: F. T. Leighton Title: Professor of Applied Mathematics