Fault-tolerant parallel algorithms

Fault-tolerant software approaches require the users to writer N versions of the algorithm which should be as different as possible to avoid correlated errors (where all or most versions fail with the same input value). The degree of difference is known as the diversity level of the versions. A new way to improve diversity is proposed. It consists of taking advantage of the higher degree of freedom provided by parallel implementations of an algorithm. Many paradigms for parallel algorithms have been developed, including divide-and-conquer, pipelined/systolic, graph function, and compute-aggregate-broadcast. By writing an algorithm using different parallel paradigms its diversity level can be increased. Software metrics can be used to quantify this level of difference