A framework for evaluating comprehensive fault resilience mechanisms in numerical programs

Abstract

As HPC systems approach Exascale, their circuit features will shrink while their overall size will grow, both at a fixed power limit. These trends imply that soft faults in electronic circuits will become an increasingly significant problem for programs that run on these systems, causing them to occasionally crash or worse, silently return incorrect results. This is motivating extensive work on program resilience to such faults, ranging from generic mechanisms such as replication or checkpoint/restart to algorithm-specific error detection and resilience mechanisms. Effective use of such mechanisms requires a detailed understanding of (1) which vulnerable parts of the program are most worth protecting and (2) the performance and resilience impact of fault resilience mechanisms on the program. This paper presents FaultTelescope, a tool that combines these two and generates actionable insights by presenting program vulnerabilities and impact of fault resilience mechanisms in an intuitive way.

FailAmp

“…A 30% overhead in sequential performance clearly sets us back a few generations in terms of Moore's law. The detailed study of resilience solutions and their overheads provided in [5] emphasizes some of these points. Another significant drawback of these detection schemes is that they have false-positive rates that are much higher than the rates at which faults themselves occur, potentially causing unnecessary recomputations.…”