Go to page

Bibliographic Metadata

 The document is publicly available on the WWW


The advance in both computation and data storage size in High Performance Computing (HPC) has not been matched by a similar advance in I/O connections. Emerging technologies have promised to overcome this gap. These could require scientific applications to change their I/O behaviour to benefit from the improvements. Therefore, a deeper analysis of applications’ I/O behaviour on modern HPC systems is required. This work defines I/O analysis criteria by which I/O behaviour can be systematically evaluated. Using the defined criteria a large set of collected I/O logs on a petascale Blue Gene/P installation, namely JUGENE, was analysed. To further the understanding of I/O architectures and their effect on I/O, a simplified parametric I/O model was developed. Results show that the implemented model has a comparable I/O behaviour to that of JUGENE, and therefore is used to evaluate new I/O technologies.