.png){kind=icon}
.png){kind=icon}
Return
Spatio-Temporal Evolving Anomaly Detection Tool for Large-Scale Heterogeneous Programs Analysis
.png){kind=icon}
PRE
AI
DOI：10.1109/tpds.2026.3711798.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Performance analysis is critical in optimizing performance. However, existing methods predominantly rely on domain expertise and focus narrowly on limited detection scopes, making program performance analysis in heterogeneous HPC systems particularly challenging under the constraints of limited expertise, architectural complexity, a huge amount of code, and large program scales. In this paper, we propose a Spatio-Temporal Evolving Anomaly Detection based tool for automatic Heterogeneous (HeSTEAD) program analysis. HeSTEAD captures spatio-temporal characteristics of programs using a novel graph representation. It further employs an optimized dynamic GNN model and an unsupervised learning method to identify anomalies within a specific execution. In addition, HeSTEAD provides hierarchical analysis that enables developers to better understand and optimize program performance across both hardware and software levels. We evaluate HeSTEAD on a real-world HPC platform with 16,000 GPUs, and the results demonstrate that HeSTEAD effectively identifies software and hardware inefficiencies in large-scale heterogeneous environments, and that the guided optimizations yield significant performance improvements.
Keywords:
Performance
heterogeneous program
anomalies
hierarchical analysis
Journal
IF:
6
Papers:
5.2K
Citations:
1.1W