We will develop a prototype I/O convolver that is a framework for combining I/O profiles that are characterizations of the rates at which a machine?s I/O and associated file systems can perform I/O operations, with I/O signatures that are records of HPC application?s file access patterns, to produce I/O performance models which can explain, predict, and be used to improve I/O performance of HPC applications. The number of external factors that impact applications performance on real systems requires additional characterization and model development to predict real world environments. Therefore despite many technical difficulties in developing truly general I/O performance models, more limited versions of useful I/O models seem achievable within the scope of an EAGER. This work is expected to have two useful outcomes: 1) predictive I/O models of two HPC applications that do significant but simple I/O for a particular file system on two architectures that can be used to explain and improve performance of these applications on the given file system 2) an experimental infrastructure that could be used to pursue the larger research agenda of developing general I/O models that scale across multiple file systems and system scales.
This project will stimulate development of new tools and approaches to deal with large scale massive file systems, their management, understanding, and measurement. The proposed modeling investigations and I/O improvements will leverage the HPC environments developed under both TeraGrid and the DoD High Performance Computing Modernization Program. The creation of I/O performance models will necessarily lead to a greater understanding of the important file system features and how they affect applications.
