Grid computing is now a viable computational paradigm for executing large scale workflow applications. However, many aspects of performance optimization remain challenging. In this paper, we focus on the workflow scheduling mechanism.  While there is much work on  static scheduling approaches for workflow applications in parallel environments, little work has been done on a real-world multi-cluster Grid environment.  Since a typical Grid environment is dynamic,  we propose a new cluster-based scheduling mechanism that dynamically executes a top-down static scheduling algorithm using the real-time feedback from the execution monitor. We also propose a novel two phase migration mechanism that mitigates the effect of a possible bad reschedule decision. Our experimental results show that this approach achieves the best performance among all the scheduling approaches we implemented on both reserved resources and those with external loads.