Fault tolerant task mapping on many-core arrays

This paper presents an approach for generating fault tolerant task mappings of applications, represented as an application process graph (APG), to a many-core array. The approach uses a multi-objective evolutionary algorithm (EA) to evolve a range of viable task mappings through the optimization of fault tolerant properties and performance criteria. Fault tolerant properties are chosen to promote task mappings that enable quick, low-cost recovery in response to fault conditions. Performance criteria promote mappings with lower network traffic. Fault tolerant properties and performance criteria tend to promote different arrangements of tasks thereby creating a range of viable mappings, represented as a Pareto front, that can be used as a pool from which a single mapping can be selected based of the prevailing demands of the system. Analysis of the evolved task maps show that they are resilient to possible fault conditions and exhibit graceful degradation of performance.