FPDetect: Efficient Reasoning About Stencil Programs Using Selective Direct Evaluation

TL;DR

FPDetect is a low-overhead method for detecting logical and soft errors in stencil computations by estimating floating-point precision bounds, enabling targeted error detection without false positives.

Contribution

It introduces an offline analysis to estimate floating-point bounds in stencil programs, allowing precise error detection with customizable accuracy and coverage.

Findings

Effective detection of logical and soft errors in stencil programs

Low overhead achieved through selective and coarse deployment of detectors

Experimental results demonstrate practicality and efficiency

Abstract

We present FPDetect, a low overhead approach for detecting logical errors and soft errors affecting stencil computations without generating false positives. We develop an offline analysis that tightly estimates the number of floating-point bits preserved across stencil applications. This estimate rigorously bounds the values expected in the data space of the computation. Violations of this bound can be attributed with certainty to errors. FPDetect helps synthesize error detectors customized for user-specified levels of accuracy and coverage. FPDetect also enables overhead reduction techniques based on deploying these detectors coarsely in space and time. Experimental evaluations demonstrate the practicality of our approach.