High-performance grid computing via distributed data access

The conventional approach to Grid computing is to utilize an additional level of user interface software ("middleware") to allow the users access the extra requirements for Grid computing in the areas of authentication, scheduling and data transport. This paradigm allows the layering of new capabilities on top of existing systems, but requires a conscious effort for their utilization. In addition many supercomputing applications have extremely large data transfer requirements and for them, this approach may be highly inefficient, or even unworkable. In this paper we describe several approaches where the data transfers associated with very large applications are addressed in a way completely transparent to the user. Each one has been implemented in an experimental mode, and where possible we provide performance numbers. Possibilities for production responsibilities are discussed.