Efficient distributed file I/O for visualization in grid environments

Large-scale simulations running in metacomputing environments face the problem of efficient file I/O. For efficiency it is desirable to write data locally, distributed across the computing environment, and then to minimize data transfer, that is, reduce remote file access. Both aspects require I/O approaches that differ from existing paradigms. For the data output of distributed simulations, one wants to use fast local parallel I/O for all participating nodes, producing a single distributed logical file, while keeping changes to the simulation code as small as possible. For reading the data file, as in postprocessing and file-based visualization, one wants to have efficient partial access to remote and distributed files, using a global naming scheme and efficient data caching, and again keeping the changes to the postprocessing code small. However, all available software solutions require all data to be staged locally (involving possible data recombination and conversion), or suffer from the performance problems of remote or distributed file systems. In this paper we show how to interface the HDF5 I/O library via its flexible Virtual File Driver layer to the Globus Data Grid. We show that combining these two toolkits in a suitable way provides us with a new I/O framework, which allows efficient, secure, distributed and parallel file I/O in a metacomputing environment.