Improving I/O performance in HPC environments using the Expand Ad-Hoc file system

This work introduces an exhaustive evaluation, including both benchmarks and real-world data-intensive applications, performed in MareNostrum 4 and HPC4AI Laboratory supercomputers using the Expand Ad-Hoc file system. Expand Ad-Hoc is an ad-hoc file system that dynamically virtualizes the local storage available on compute nodes (i.e., SSD, SHM, etc.) into a fast storage volume to reduce congestion on parallel file systems used as backends in High-Performance Computing (HPC) environments. The main contributions of this work include the design of a new ad-hoc parallel file system, called Expand Ad-Hoc, compatible with POSIX and MPI-IO, and an exhaustive and comprehensive evaluation of Expand. The evaluation compares the performance obtained using IOR and DLIO benchmarks and Nek5000 and Remote Sensing real-world applications on Expand Ad-Hoc, GekkoFS, GPFS, and BeeGFS, showing satisfactory results proving that Expand Ad-Hoc can be used in HPC environments to improve the I/O performance of data-intensive applications transparently.