Improving Scaling in Conquest for HECToR Phase 3 - OpenMP-MPI Hybrid Implementation
This distributed Computational Science and Engineering (dCSE) project is aimed to improve scaling properties of the ab initio Density Functional Theory (DFT) code Conquest running on HECToR using the hybrid OpenMP-MPI parallelisation implementation. This work will enhance the ability of the researchers in the electronic structures community to do accurate ab initio quantum mechanical simulations on a very large scale (over a million atoms) while using modest computational resources.
The key aims of the project were to:
- Develop an initial OpenMP implementation of Conquest and characterise the communications patterns for different systems.
- Develop a new MPI approach to fit better with OpenMP.
- Tune the new hybrid OpenMP / MPI approach.
The individual achievements of the project are summarised below:
- Conquest code has been profiled on the current architecture of HECToR (Phase3).
- The bottle-necks and hottest part of the code have been identified to be in the matrix multiplication subroutines.
- OpenMP parallelisation has been successfully implemented in the matrix multiplication subroutines.
- Extensive tests have been done and results show that the hybrid OpenMP-MPI implementation does significantly improve the weak-scaling property of the code on HECToR when compared with the original pure MPI implementation.
- Linear weak-scaling has been restored to Conquest when running on HECToR.
- Several versions of the OpenMP parallelisation methods have been implemented and compared. These include an implementation which uses explicit tasks.
Please see PDF or HTML for a report which summarises this project.
