מושגי ליבה
This paper introduces a novel approach using embedding and deep learning graph neural networks (GNNs) to identify bugs in MPI programs, achieving high accuracy in detecting error types. The models trained on benchmark suites show promising results for error detection.
תקציר
The paper presents a pioneering method utilizing embedding and GNNs to detect errors in MPI programs, achieving high accuracy. Training on benchmark suites MBI and MPI-CorrBench, the models demonstrated effectiveness in identifying various error types with over 80% accuracy. The study highlights the importance of feature selection, normalization strategies, and compiler optimization for improving prediction accuracy. Additionally, an ablation study reveals insights into error interaction patterns and model generalization capabilities across different datasets.
Key points:
Introduction of novel approach using embedding and GNNs for error detection in MPI programs.
High accuracy achieved in detecting various error types through training on MBI and MPI-CorrBench.
Importance of feature selection, normalization strategies, and compiler optimization for improved prediction accuracy.
Ablation study provides insights into error interaction patterns and model generalization capabilities.
סטטיסטיקה
By training our models on the same benchmark suite, we achieved a prediction accuracy of 92% in detecting error types.
We achieved a promising accuracy of over 80% when transitioning from MBI to MPI-CorrBench.
The detection accuracy of removed errors varied significantly between 20% to 80%, indicating connected error patterns.