Workshop: Workshop on Operational Data Analytics in HPC (HPC-ODA)
Authors: Kazuki Miyake (Kobe University, RIKEN R-CCS); Naohisa Sakamoto (Kobe University); Jorji Nonaka, Masaaki Terai, and Yuji Iguchi (RIKEN R-CCS); and Takanori Fujiwara (The University of Arizona)
Abstract: In this paper, we apply Tensor Unified Linear Comparative Analysis (TULCA) to rack-level power consumption data from the supercomputer Fugaku. The data are represented as a third-order tensor with temporal, spatial, and variable axes, with each time point labeled by its operational period. TULCA generates low-dimensional representations that emphasize differences among labeled groups while preserving the tensor structure. Based on this representation, we construct an analysis workflow that provides an overview of temporal behavior and traces differences between analyst-selected temporal clusters back to the contributing compute racks. We applied the workflow to power consumption data from Fugaku's 432 compute racks across two operational periods. In case studies, the racks contributing most strongly to the observed separation exhibited sharp, short-term power fluctuations. These results demonstrate the potential of the workflow for interactive comparative analysis of HPC operational data.
Back to Workshop on Operational Data Analytics in HPC (HPC-ODA) Archive Listing
Back to Full Workshop Archive Listing