Poster: Reservation-Based I/O Performance Guarantee for MPI-IO Applications Using Shared Storage Systems

摘要

While optimized collective I/O methods are proposed for MPI-IO implementations, a problem in concurrent use of the shared storage system is raised. In order to prevent performance degradation of parallel I/O due to such I/O conflict, we propose an advance reservation approach, including possible integration with existing batch scheduler on HPC clusters. In this work, we use Dynamic-CoMPI as a MPI-IO implementation and Papio as a shared storage system which implements parallel I/O and performance reservation. Then we have been developing the ADIO layer to connect these systems and to evaluate the benefits of the reservation-based performance isolation. Our prototype implementation, Dynamic-CoMPI/Papio, was evaluated using the MPI-IO Test benchmark and the BISP3D application. In our preliminary evaluation, total execution time increased 3~12% with the Dynamic-CoMPI/PVFS2 and 6~40% with the Dynamic-CoMPI/Lustre, under the situation where additional workload affected on MPI execution, however there was no obvious time increase with Dynamic-CoMPI/Papio.