DETOx: Towards Optimal Software-based Soft-Error Detector Configurations

TL;DR

This paper presents DETOx, a method to optimize the selection of assertions in software to minimize silent data corruptions caused by hardware faults, balancing detection effectiveness and runtime overhead.

Contribution

It introduces a novel approach to identify the optimal subset of assertions that reduces silent data corruptions without extensive fault-injection testing.

Findings

Effective assertion subset selection reduces SDCs

Method balances detection and runtime overhead

No need for exhaustive fault-injection experiments

Abstract

Application developers often place executable assertions -- equipped with program-specific predicates -- in their system, targeting programming errors. However, these detectors can detect data errors resulting from transient hardware faults in main memory as well. But while an assertion reduces silent data corruptions (SDCs) in the program state they check, they add runtime to the target program that increases the attack surface for the remaining state. This article outlines an approach to find an optimal subset of assertions that minimizes the SDC count, without the need to run fault-injection experiments for every possible assertion subset.