Error Resilience of Three GMRES Implementations under Fault Injection

Abstract

The resilience behavior of three GMRES prototyped implementations (with Incomplete

LU, Flexible and randomized-SVD—based preconditioners) has been analyzed

with a soft errors injection approach. A low-level fault injector is inserted into the

GMRES solvers, which randomly select locations in the program to inject the fault

across multiple executions. This fault injection approach combines the configurability

of high-level and the accuracy of low-level techniques at the same time, so the

effect of faults may be closely emulated. In order to gather enough statistical data, a

set of eighteen sparse matrix-based linear systems Ax = b has been solved with these

GMRES implementations in the injection experiments and monitored. The results of

this prototype-based fault injection suggest an improved error resilience behavior of

the randomized-SVD—based preconditioned GMRES version in many of the analyzed

matrices, which points out to its interest in supercomputing applications where

silent errors are more prominent