Applications Classification and Scheduling on Heterogeneous HPC Systems Using Experimental Research

Heterogeneous high-performance computing (HPC) systems have been proposed as a power efficient alternative to traditional homogeneous systems. In heterogeneous HPC system, fast CPUs which have complex pipelines, high clock frequencies as well as high power consumption are combined with slow ones which have simple pipelines, low clock frequencies as well as low power consumption. Different types of applications should be distributed to run on different type of CPUs, in order to achieve a trade-off between energy and performance. In this paper, we use experimental research to investigate the applications classifications which categorize each application as the CPU-intensive, memory-intensive, or phase-change application. We also conduct some experiments to measure the current, power, and energy of different types of applications. Afterwards, a scheduling method for applications on heterogeneous HPC systems is proposed. The experiment shows that the scheduling method can trade off the latency and energy consumption based on the applications classification and measurement.