Towards Real- Time Classification of HPC Workloads via Out-of-band Telemetry

Detecting illicit workloads on High Performance Computing (HPC) systems is an important task. Such workloads might indicate a user account is being misused, for example to run cryptocurrency miners or password crackers. Existing solutions use in-band collection of data, which may slow down the workloads on the system. We present early results demonstrating that real-time classification of HPC workloads based on out-of-band telemetry may be possible. Further, we show a method-dependent accuracy of greater than 99 %, with the best results achieving near-perfect classification. These results suggest that creating a real-time classifier of HPC jobs that does not impact system performance is possible.