On Being Close to the Data: Executing Code in a Replica Management System

The goal of scientific grid computing is to enable researchers to attack problems that were previously impossible. Cooperative storage systems have been assembled to provide users with access to surprisingly large amounts of useful space, but locating, accessing, and efficiently employing such distributed data sets in scientific simulation or analysis is another monumental problem. In this paper, we describe simple techniques compatible with existing software that may be used on clusters or grids to access catalogued, replicated data sets: software tools that allow user code to use a replica system as a scientific file system. The importance of operating in-place, close to the data, is demonstrated by the performance improvements gained by using data locations as a guide when scheduling computation. A new replica management system, called GEMS1, is proposed to improve the research capabilities of two important computing infrastructures: university networks (groups of clusters) and volunteer computer resources (Internet computing). I