Profiling and Modeling of Power Characteristics of Leadership-Scale HPC System Workloads

TL;DR

The paper tackles the challenge of profiling and modeling power consumption for leadership-scale HPC workloads to enable energy-aware operation. It proposes a low-latency ML pipeline that converts high-resolution power telemetry into fixed-length feature representations via a 186-feature extractor and a GAN-based latent space, followed by DBSCAN clustering into 119 contextualized classes across ~60K Summit jobs. An open-set CAC-based classifier handles unseen patterns, and an iterative workflow with periodic retraining maintains accuracy as workloads evolve. The approach yields actionable insights into the power landscape across science domains and demonstrates robust performance in both closed-set and open-set settings, offering a path toward energy-aware resource management in exascale systems.

Abstract

In the exascale era in which application behavior has large power & energy footprints, per-application job-level awareness of such impression is crucial in taking steps towards achieving efficiency goals beyond performance, such as energy efficiency, and sustainability. To achieve these goals, we have developed a novel low-latency job power profiling machine learning pipeline that can group job-level power profiles based on their shapes as they complete. This pipeline leverages a comprehensive feature extraction and clustering pipeline powered by a generative adversarial network (GAN) model to handle the feature-rich time series of job-level power measurements. The output is then used to train a classification model that can predict whether an incoming job power profile is similar to a known group of profiles or is completely new. With extensive evaluations, we demonstrate the effectiveness of each component in our pipeline. Also, we provide a preliminary analysis of the resulting clusters that depict the power profile landscape of the Summit supercomputer from more than 60K jobs sampled from the year 2021.

Figures (10)

• Figure 1: End-to-end flow diagram of the data-driven analysis and classification modeling pipeline.
• Figure 2: Timeseries of typical HPC workloads. The subfigures help visualize that the features calculated in Table \ref{['tab:data_feature_spec']} can distinguish between different power profiles. Background shades in each subplot correspond to 4 bins of timeseries.
• Figure 3: GAN model for latent feature generation. It takes input of dimension $\mathbb{R}_{x}$ and generate the latent feature of lower dimension $\mathbb{R}_{z}$. Once trained, the model can generate latent features for new data points.
• Figure 4: Data distribution of three reconstructed and real features showing that the reconstructed distribution is similar to the real distribution. Reconstruction of original features confirm that the low-dimension latent feature have necessary information of the data-points as in the high-dimension space.
• Figure 5: Groupings of power profiles based on the power utilization patterns exhibited by jobs. The different classes can be grouped into high-level classification into three types. Classes from 0-20 are compute intensive, 21-92 are mixed type, and 93-118 are non-compute intensive jobs. Tiles with blue curve have high power consumption and plots with green curve have low power consumption. Intensity of the background color of each tile represents the density of each pattern among the entire population.