Efficient Workload Balancing on Heterogeneous GPUs using Mixed-Integer Non-Linear Programming

Recently, heterogeneous system architectures are becoming mainstream for achieving high performance and power efficiency. In particular, many-core graphics processing units (GPUs) now play an important role for computing in heterogeneous architectures. However, for application designers, computational workload still needs to be distributed to heterogeneous GPUs manually and remains inefficient. In this paper, we propose a mixed integer non-linear programming (MINLP) based method for efficient workload distribution on heterogeneous GPUs by considering asymmetric capabilities of GPUs for various applications. Compared to the previous methods, the experimental results show that our proposed method improves performance and balance up to 33% and 116%, respectively. Moreover, our method only requires a few overhead while achieving high performance and load balancing.