Energy implications of desynchronization-based runtime improvements

Determine whether runtime improvements obtained through desynchronization translate into energy savings, using energy-to-solution analysis based on per-node power measurements such as LIKWID’s RAPL-based energy counters.

Background

The paper demonstrates that beneficial desynchronization between compute-bound and memory-bound kernels can substantially improve runtime performance, including in the Parallel Optical Flow Solver and the Pisolver-Triad benchmark. However, faster execution does not necessarily imply lower energy consumption, particularly when desynchronization changes memory-bandwidth utilization, processor activity, or the duration and intensity of resource contention.

The authors explicitly identify energy-to-solution as a future investigation and propose per-node power measurements using LIKWID’s RAPL-based energy counters to assess whether the observed runtime benefits also reduce total energy consumption. The question remains unresolved in the paper because no energy measurements or conclusions are reported.