High Performance Software Reconfiguration in the Context of Distributed Systems and Interconnection Networks.
Date of Award
Doctor of Philosophy (PhD)
S. Sitharama Iyengar
Designed algorithms that are useful for developing protocols and supporting tools for fault tolerance, dynamic load balancing, and distributing monitoring in loosely coupled multi-processor systems. Four efficient algorithms are developed to learn network topology and reconfigure distributed application programs in execution using the available tools for replication and process migration. The first algorithm provides techniques for transparent software reconfiguration based on process migration in the context of quadtree embeddings in Hypercubes. Our novel approach provides efficient reconfiguration for some classes of faults that may be identified easily. We provide a theoretical characterization to use graph matching, quadratic assignment, and a variety of branch and bound techniques to recover from general faults at run-time and maintain load balance. The second algorithm provides distributed recognition of articulation points, biconnected components, and bridges. Since the removal of an articulation point disconnects the network, knowledge about it may be used for selective replication. We have obtained the most efficient distributed algorithms with linear message complexity for the recognition of these properties. The third algorithm is an optimal linear message complexity distributed solution for recognizing graph planarity which is one of the most celebrated problems in graph theory and algorithm design. Recently, efficient shortest path algorithms are developed for planar graphs whose efficient recognition itself was left open. Our algorithm also leads to designing efficient distributed algorithm to recognize outer-planar graphs with applications in Hamiltonian path, shortest path routing and graph coloring. It is shown that efficient routing of information and distributing the stack needed for for planarity testing permit local computations leading to an efficient distributed algorithm. The fourth algorithm provides software redundancy techniques to provide fault tolerance to program structures. We consider the problem of mapping replicated program structures to provide efficient communication between modules in multiple replicas. We have obtained an optimal mapping of 2-replicated binary trees into hypercubes. For replication numbers greater than two, we provide efficient heuristic simulation results to provide efficient support for both 'N-version programming' and 'Recovery block' approaches for software replication.
Hegde, Vinayak Gajanan, "High Performance Software Reconfiguration in the Context of Distributed Systems and Interconnection Networks." (1993). LSU Historical Dissertations and Theses. 5510.