Abstract
The ability to effectively use computational grids for data intensive applications is becoming increasingly important. The distributed, heterogeneous, shared nature of the computing resources provides a significant challenge in developing support for computationally demanding applications. In this paper we describe several performance optimization techniques we have developed for the filter-stream programming framework that we have designed and implemented for programming data intensive applications on the Grid. We present performance results for multiple versions of a medical imaging application on various distributed machine configurations that show the benefits of the optimizations, and also provide evidence that filter-stream programming can be implemented to both efficiently utilize available Grid resources and to provide scalable application performance as additional resources are made available.