Numerical Defect Correction as an Algorithm-Based Fault Tolerance Technique for Iterative Solvers

As hardware devices like processor cores and memory sub-systems based on nano-scale technology nodes become more unreliable, the need for fault tolerant numerical computing engines, as used in many critical applications with long computation/mission times, is becoming pronounced. In this paper, we present an Algorithm-based Fault Tolerance (ABFT) scheme for an iterative linear solver engine based on the Conjugated Gradient method (CG) by taking the advantage of numerical defect correction. This method is "pay as you go", meaning that there is practically only a runtime overhead if errors occur and a correction is performed. Our experimental comparison with software-based Triple Modular Redundancy (TMR) clearly shows the runtime benefit of the proposed approach, good fault tolerance and no occurrence of silent data corruption.