Efficient, large-scale Density Functional Theory on the cloud

As part of project Quantifi, an Innovate UK funded project, Rahko, Johnson Matthey and Amazon Web Services (AWS) have jointly demonstrated the use of high-performance computing (‘HPC’) to deliver efficient, large-scale Density Functional Theory (‘DFT’) calculations on the cloud. 


Accelerating the discovery of new chemicals and materials 

Discovery of new chemicals and materials has historically been a slow and expensive process due to the enormous associated search space and high experimental costs. 

In recent years, computational methods have ushered in a new paradigm in the discovery process that allows for a steadily decreasing reliance on extensive laboratory experiments, accelerating the discovery process and significantly reducing associated costs. 

These computational methods are critically enabled by HPC infrastructure. 


DFT with HPC

DFT is a computationally efficient quantum mechanical approach that enables the accurate non-empirical modelling of many unexplored regions of the chemical space. 

The evolution of HPC is an integral part in the growth and application of DFT methods. DFT methods have over the last two decades have become a standard tool in the materials development workflow. Using these quantum mechanical methods allows us to extend the bounds of searchable space and increase the complexity of the systems that can be computationally handled. 

Cloud computing is a relatively new addition to the HPC toolset that enables flexible, on-demand access to large-scale compute-intensive DFT calculations. 


Scaling calculations on the cloud

Cloud computing provider AWS offers parallel clusters with a low latency Elastic Fabric Adapter (‘EFA’) suitable for scaling calculations up to thousands of nodes with fast internode communication.  

The demonstration of scaling shows that for the chemical systems tested, the EFA was able to match conventional low latency technologies. 

Quantifi researchers’ test calculations on a Pd42O42 cluster and a Pd27Ag4Pt alloy demonstrate that 98-99% parallelization can be achieved on multiple c5n.18xlarge nodes for large gas phase cluster and solid state calculations with the Quantum Espresso materials modelling software using the Intel Math Kernel Library, and Intel MPI as can be seen from the figure on strong scaling results below. 

The estimated speedup for the systems investigated is up to 50-100x, with speedups expected to increase in line with larger system sizes. 

This implementation has been a valuable addition to Hyrax, Rahko’s quantum discovery platform, and is easily integrated into existing discovery workflows. 

Figure Strong scaling results comparing large cluster and alloy calculations on AWS ParallelCluster