GPUs (Graphics Processing Units) have gained wide adoption as accelerators for general purpose computing. They are widely used in error-sensitive applications, i.e. General Purpose GPU (GPGPU) applications However, the reliability implications of using GPUs are unclear. This paper presents a fault injection study to investigate the end-to-end reliability characteristics of GPGPU applications. The investigation showed that 8% to 40% of the faults result in Silent Data Corruption (SDC). To reduce the percentage of SDCs, we propose heuristics to selectively protect specific elements of the application and design error detectors based on heuristics. We evaluate the efficacy of the detectors in reducing SDCs and measure performance overheads of the detectors. Our results show that the heuristics are able to reduce the SDC causing faults by 60% on average, while incurring reasonable performance overheads (35% to 95%).
email: email@example.comPhone: 604-827-4245 (please email first)
Address: Rm. 4048, Fred Kaiser Building, 2332 Main Mall, Vancouver, BC V6T1Z4.
- An Empirical Study of the Impact of Single and Multiple Bit-Flip Errors in Programs
- New Wine in an Old Bottle: N-Version Programming for Machine Learning Components
- Stealthy Attacks Against Robotic Vehicles Protected by Control-based Intrusion Detection Techniques
- How Far Have We Come in Detecting Anomalies in Distributed Systems? An Empirical Study with a Statement-level Fault Injection Method
- TensorFI: A Flexible Fault Injection Framework for TensorFlow Applications
- GPU-TRIDENT: Efficient Modeling of Error Propagation in GPU Programs
- How Effective are Smart Contract Static Analysis Tools ? Evaluating Smart Contract Static Analysis Tools Using Bug Injection
- TraceSanitizer – Eliminating the Effects of Non-determinism on Error Propagation Analysis
- Improving the Accuracy of IR-Level Fault Injection
- Out of Control: Stealthy Attacks on Robotic Vehicles Protected by Control-Based Techniques