Workshops Archive

Quantifying Workload Exposure in HPC Power Telemetry at Operational Sampling Rates


Workshop: 5th International Workshop on Cyber Security in High Performance Computing

Authors: Kwesi Danso, Samuel Abiola, and Susan Mengel (Texas Tech University)

Abstract: HPC centers publish power telemetry to support monitoring, scheduling, and optimization. While this telemetry is published for operational use, it can also reveal information about the workloads on a system. Prior work shows workloads leave signatures in high-frequency telemetry. How much of that signature survives at the coarse rates used in production is unknown. We study workloads using CPU power telemetry from Intel RAPL at approximately 1 kHz, downsampled to intervals as coarse as 60 s, measuring recoverability for resource classes and for individual workloads on unseen problem sizes. Recoverability declines as sampling coarsens, from 0.87 to 0.71 for resource class and 0.74 to 0.54 for workload identity. In short jobs, temporal information distinguishing individual workloads disappears under coarse sampling, while the steady power composition remains. At coarse rates, distinct workloads remain identifiable while similar ones become indistinguishable. Thus, coarsening the sampling rate is a practical but incomplete privacy control.


Back to 5th International Workshop on Cyber Security in High Performance Computing Archive Listing

Back to Full Workshop Archive Listing