A gracefully degradable VLSI system for linear programming

The use of a fault-tolerant VLSI system for storing and solving linear programming problems is presented. The system can bear multiple faults in processing elements and/or links and still function with an acceptable performance degradation. It is based on an interconnection pattern consisting of a complete binary tree in which spare links between cousin nodes are added so as to reconfigure it as a ternary tree. At any given time of a computation, faulty processing elements and/or links are circumvented by using such spare links. It is shown that the total silicon area required by this structure is only a constant factor higher than that of a complete binary tree. The result is used to give an efficient implementation of the simplex algorithm in which the time required to perform a single pivot step matches a previously established lower bound for tree machines in spite of faults.< >