SC17 Denver, CO

P84: PRESAGE: Selective Low Overhead Error Amplification for Easy Detection


Authors: VIshal Chandra Sharma (University of Utah), Arnab Das (University of Utah), Ian Briggs (University of Utah), Ganesh Gopalakrishnan (University of Utah), Sriram Krishnamoorthy (Pacific Northwest National Laboratory)

Abstract: Soft-errors remain a vexing challenge. Today's error detectors either come with high false positives or high omissions. Our work focuses on not losing an error sown, but amplifying it so that cheap detection is enabled.

Consider structured address generation in loops where data from a base address plus offset is used. An erroneous offset no longer crashes today's applications thanks to large address spaces; instead, it silently corrupts data (unintended fetch). We relativize address generation using LLVM, thus each new address is not base plus offset but previous relative address plus offset. If one address is corrupted, all future address are corrupted in a chain. This permits efficient loop exit-point detection.

Relativization has low overhead, actually lowered by some ISAs down to zero. These advantages survive crucial compiler memory access optimizations. We demonstrate 100% SDC detection for a class of benchmarks with respect to structured address protection.

Award: Best Poster Finalist (BP): no

Poster: pdf
Two-page extended abstract: pdf


Poster Index