ARcode: HPC Application Recognition Through Image-encoded Monitoring Data

TL;DR

ARcode is a novel method that encodes HPC job monitoring data into images and uses CNNs for application recognition, achieving higher accuracy than existing methods.

Contribution

This paper introduces ARcode, a new image-encoding approach combined with CNNs for automatic feature learning in HPC application recognition.

Findings

ARcode outperforms state-of-the-art methods by up to 18.87% in accuracy.

ARcode achieves over 20% improvement for specific applications at high confidence.

The approach effectively leverages image encoding and deep learning for system management.

Abstract

Knowing HPC applications of jobs and analyzing their performance behavior play important roles in system management and optimizations. The existing approaches detect and identify HPC applications through machine learning models. However, these approaches rely heavily on the manually extracted features from resource utilization data to achieve high prediction accuracy. In this study, we propose an innovative application recognition method, ARcode, which encodes job monitoring data into images and leverages the automatic feature learning capability of convolutional neural networks to detect and identify applications. Our extensive evaluations based on the dataset collected from a large-scale production HPC system show that ARcode outperforms the state-of-the-art methodology by up to 18.87% in terms of accuracy at high confidence thresholds. For some specific applications (BerkeleyGW and e3sm), ARcode outperforms by over 20% at a confidence threshold of 0.8.