Self-Adaptive Heterogeneous Cluster with Wireless Network

The high performance computing (HPC) community has been exploring novel platforms to push performance forward. Field Programmable Logic Arrays (FPGAs) and Graphics Processing Units (GPUs) have been widely used as accelerators for computational intensive applications. Heterogeneous cluster is one of the promising platforms as it combines characteristics of multiple processing elements, to meet requirements of various applications. In this work, we build a self-adaptive framework for heterogeneous clusters, coupled with a customised wireless network. A runtime cluster model is implemented to predict throughput, power and thermal merits for heterogeneous clusters. Cluster configurations are scheduled to improve cluster power efficiency, as well as to reduce peak temperature of processing elements. Results show that, for monitoring operations upon heterogeneous clusters, the customised wireless network provides stable and scalable performance for negligible overhead. A high performance application is developed under the proposed framework. Experiments show that this approach can improve both power efficiency and energy efficiency of N-body simulation for more than 15 times, while reducing device peak temperature by up to 12° C.