A Scalable Framework for Heterogeneous GPU-Based Clusters