Facilitating Electronic Structure Computations on GPU-based Exascale Platforms

Explore the challenges and solutions for porting electronic structure codes to exascale platforms in this webinar from the Exascale Computing Project. Dive into the implementation strategies and algorithmic choices made in the PROGRESS and BML libraries developed by ECP's CoPA project. Learn about performance portability, strong scaling, and GPU offloading techniques using OpenMP. Discover how to speed up electronic structure calculations and run molecular dynamics simulations on exascale platforms. Examine the main numerical kernels for electronic structure calculations and the development of alternative solvers based on polynomial matrices. Gain insights into computer science challenges, matrix formats, and the importance of Fortran interfaces for targeted application codes. Explore benchmarking strategies, performance optimization on GPU-based systems, and the use of Chebyshev expansions for density matrix calculations. Understand the balance between computational cost and accuracy through matrix thresholding, and examine parallel scaling performance on Summit supercomputer. Conclude with valuable lessons learned about efficiently utilizing GPUs for electronic structure computations on exascale platforms.

HPC Best Practices Webinar Series
Algorithms and performance portability for electroni structure
Speeding up electronic structure calculations to ena
Running MD on exascale platforms
Main numerical kernels for electronic structure calculations
Developing alternative solvers based on polynomial matrices
Implementation divided into two libraries
Using OpenMP for GPU offloading
General implementation strategy
Computer Science challenges
BML: supported (shared memory) matrix formats
BML: Supporting multiple data types in a C code
BML: Fortran interface is important for targeted application codes
BML: Unit test/Continuous integration
Offloading to GPU
Offloading strategy
GPU offloading with OpenMP
Challenges in interfacing with optimized vendor libra
Using a synthetic Hamiltonian matrix for Performanc Benchmarking
rocSPARSE performance on Crusher @ OLCF
Chebyshev expansions for modest matrix sizes (metals)
Chebyshev expansion of Density Matrix
Exploiting GPU concurrency in calculating Chebysh terms
Distributing computation
Balancing computational cost and accuracy with matrix thresholding
A non-intrusive implementation
What about wavefunction-based solver? (Planewaves...)
Numerical Discretization of DFT problem
Parallel scaling/performance on Summit
Lesson learned: Efficiently using GPUs requires a lo work!
Acknowledgments