Computational grids are some of the largest computer systems in existence today. Unfortunately they are also, in many cases, the least reliable. This research examines the use of redundancy with permutation as a method of improving reliability in computational grid applications. Three primary avenues are explored -- development of a new redundancy model, the Replication and Permutation Paradigm (RPP) for computational grids, development of grid simulation software for testing RPP against other redundancy methods and, finally, running a program on a live grid using RPP. An important part of RPP involves distributing data and tasks across the grid in Latin Square fashion. Two theorems and subsequent proofs regarding Latin Squares are developed. The theorems describe the changing position of symbols between the rows of a standard Latin Square. When a symbol is missing because a column is removed the theorems provide a basis for determining the next row and column where the missing symbol can be found. Interesting in their own right, the theorems have implications for redundancy.;In terms of the redundancy model, the theorems allow one to state the maximum makespan in the face of missing computational hosts when using Latin Square redundancy. The simulator software was developed and used to compare different data and task distribution schemes on a simulated grid. The software clearly showed the advantage of running RPP, which resulted in faster completion times in the face of computational host failures. The Latin Square method also fails gracefully in that jobs complete with massive node failure while increasing makespan. Finally an Inductive Logic Program (ILP) for pharmacophore search was executed, using a Latin Square redundancy methodology, on a Condor grid in the Dahlem Lab at the University of Louisville Speed School of Engineering. All jobs completed, even in the face of large numbers of randomly generated computational host failures. vi
展开▼
