Technical Papers Archive

BitIR: Cross-Architecture Fault Injection for Resilience Analysis of Heterogeneous GPU Applications


Authors: Maisy Dunlavy and Zhiling Lan (University of Illinois Chicago) and Michael E. Papka (University of Illinois Chicago, Argonne National Laboratory)

Abstract: Modern GPU-based HPC systems rely on heterogeneous vendor stacks, yet reliability studies are largely limited to single architectures. Consequently, it remains unclear how faults behave across different GPU backends. We present BitIR, a cross-architecture fault injection framework that enables semantically aligned, software-visible resilience analysis across NVIDIA, Intel, and AMD GPUs. BitIR injects deterministic faults at the LLVM IR level, ensuring equivalent bit perturbations prior to backend lowering and enabling direct cross-vendor comparison. Using a set of representative GPU benchmarks, we conduct large-scale injection campaigns on production systems and classify outcomes into masked results, silent data corruption (SDC), failures, and hangs. We find that identical faults produce markedly different behaviors across vendors: Intel most often masks faults but exhibits more hangs when faults escape masking; NVIDIA exposes more overt detectable failures; and AMD alternates between SDC-dominant and failure-dominant behavior depending on the benchmark and fault site.


Presentation: file

Back to Technical Papers Archive Listing