Exploring Fine-Grained Task-based Execution on Multi-GPU Systems