Performance Prediction under Power Capping

High-performance computing (HPC) clusters have been constantly increasing in size as well as in power consumption. In the future, these clusters will likely be power capped since the power supply is limited by the surrounding infrastructure. Therefore, the peak power draw of running jobs can not be guaranteed which in turn increases jobs' execution time and reduces the cluster throughput. On the other hand, jobs have distinct power draws and their performance scales differently under power constraints. It is vital to understand power and performance behaviors, in order to consume the limited power budget effectively. We propose in this work a model that predicts performance of power capped applications. Applying this model, comprehensive explorations with distinguished power settings can be avoided. The model has been validated as accurate: predicted performance differs barely from the measured average, below 3%.