Efficient Parallel I/O in HDF5 for Accelerator Computing
Dr. Li and his student will also work with us in optimizing performance and scalability of HDF5 and I/O operations at NC State University.
This project aims to improve the efficiency of moving data between multiple GPUs and the parallel file system for upcoming exascale supercomputers. NC State researchers Dr. Michela Becci and Dr. Li and their students will profile existing I/O benchmarks and optimize HDF5 and I/O performance. Designs for asynchronous I/O for GPUs and node-local storage on a compute node will be updated to maximize performance and scalability. This research will help ensure efficient parallel I/O in accelerated computing nodes and enable the success of exascale computing.
Sponsor
Lawrence Berkeley National Laboratory - University of California - Berkeley
The grant—running from August 17, 2020 to October 31, 2022—is for a total of $114,453.
Principle Investigators
Michela Becchi
More Details
HDF5 is designed to store and manage high-volume and complex science data and has become the leading I/O middleware solution at DOE supercomputing centers. As upcoming exascale supercomputers are using accelerators, such as graphical processing units (GPUs), for improving the performance of computing, data must be moved efficiently between storage and accelerators. To perform efficient parallel I/O in accelerated computing nodes for moving data between multiple GPUs and the parallel file system using node-local storage devices and network interconnects, this project will extend asynchronous I/O. In this project, Dr. Michela Becci and her student will work with us in identifying I/O benchmarks representative of ECP applications and profiling their current performance. The project will then update the designs of asynchronous I/O for using GPUs and node-local storage on a compute node.