A centralized data access model for grid computing

Global access to storage is a common theme of grid computing, with access mechanisms often enforcing a major restriction on the distribution of significant applications across a computational grid. The established approach is to distribute the data with the jobs, sometimes requiring lengthy delays on job completion and the necessity for significant resource discovery to establish local data capabilities. For applications that are truly data intensive, this may render them either highly inefficient, or even incapable of using grid computing environments. In this paper we describe a different approach, where the opportunity to design a grid environment from scratch is used to build a tightly-coupled, data-oriented infrastructure that leverages deep investment in leading edge technology to provide very high-speed, widespread access to large data storage. Results from a geographically distributed grid established for the Supercomputing 2002 conference, using preliminary TeraGrid infrastructure, are included and show encouraging performance including data transfer rates of over 700 MB/s using eight 1 Gb/s links from a storage area network to a 10 Gb/s wide area network.