Prior work has shown that power consumption
traces of HPC workloads exhibit distinctive statistical characteristics,
which allows the workload that generated a given power
trace to be inferred with high accuracy. However, these power
signatures apply to the entire power trace, with no ability to break
it down further into phases or to recognize novel combinations
of known workloads.
In this work, we propose and evaluate a mechanism for
partitioning a power trace into phases and matching each phase
to a known kernel or workload. We evaluate this technique on a
set of 388 power traces collected from 21 benchmarks, including
CPU-intensive system stressors; the NAS Parallel Benchmarks;
and Mahout data analytics workloads. Our technique is able to,
on average, attribute 78% of the points in a concatenated trace
to the correct kernel.