BBView: A View-Aware Burst-Buffer Mechanism for MPI-IO

Large-scale simulations periodically enter an output phase—for example to save a restart image or dump a multidimensional field for post-processing. Because each MPI process owns a scattered subset of the global grid, these phases issue millions of tiny, non-contiguous writes whose transfer sizes are far too small to exploit the peak bandwidth of modern NVMe drives and parallel file systems.We present BBView, a lightweight Open MPI plug-in that buffers the MPI file view held by each process. Instead of scattering a rank’s contiguous in-memory data across the many disjoint offsets prescribed by its view, BBView defers that scatter—it first writes the data out exactly as it sits in memory, in one contiguous stream to a node-local file. The moment the process later closes the file (or installs a new view), a background daemon reconstructs the stored view and scatters the data to the parallel file system, ensuring the shared file is identical to the layout the application specified.BBView is seamlessly compatible with applications built on HDF5 or PnetCDF, and functions as an effective I/O acceleration mechanism. An evaluation on 128 nodes showing up to 74.2 × higher throughput than UnifyFS on local NVMe SSD, driven purely by larger transfer sizes.