HPC IO Analysis is a community effort aiming at fostering collaboration and improving knowledge of the I/O aspects in HPC applications. Our study aims at creating a shared and open-access database of the I/O performance of applications running on different parallel I/O libraries and in different layers of the I/O stack.
Dealing with large HPC applications on modern infrastructures has become a challenge for most of organizations. This is especially acute for the I/O stack. However, the complexity and heterogeneity of today's scientific and HPC applications make this study challenging. The challenge arises from the different hardware availability and the applications-specific instrumentation efforts; therefore, existing analyses are limited to a few hardware settings and a specific family of applications. An in-depth understanding of I/O characteristics and requirements of different varieties of HPC applications, such as metadata operations and I/O access patterns, is crucial for designing efficient storage and I/O solutions.
The study's comprehensiveness can benefit the whole community of researchers to analyze and spot I/O bottlenecks and further design more efficient system software.
HPC IO Analysis propose to upload I/O traces and will provide in return an analysis. Both the trace and the analysis are made publicly available. Thus, all uploaded data will have to comply with the Creative Common License.
Refer to our wiki for an introduction on getting the traces.
Application | Workload Family | Institute | Cluster/TOP500 | IO Library | Checkpointing | MPI Ranks | Actions |
---|---|---|---|---|---|---|---|