Section: New Results

New algorithm for I/O scheduling

We started working on I/O scheduling for HPC applications. HPC applications can be caracterized by I/O patterns that are repeated periodically. We showed in a simple context how this information can be taken into account to outperform state of the art I/O schedulers [15].

These preliminary results led to the obtention of the ANR Dash (see Section 9.1.2).

After which, we have performed a theoretical analysis to show how one should size the burst-buffers and the bandwidth to those buffers on a HPC system depending on the applications running. In our study we focused on one role of the buffers (namely the role of buffer to the PFS) [42]. This study is particularly important since those buffers are limited and can be used for many usage. Over or under booking the buffers for a specific use leads to an increase of congestion.