Multi-node parallelization performance of first-principles calculations
Quantum ESPRESSO (QE) is a computer simulation package based on Density Functional Theory (DFT) for calculating electronic and structural properties of a material at ground state, which gives an excellent balance of accuracy and computational cost. For a macromolecular system with a large number of...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2019
|
| Subjects: | |
| Online Access: | http://eprints.utm.my/id/eprint/102489/1/MohammedBubaMFS2019.pdf.pdf http://eprints.utm.my/id/eprint/102489/ http://dms.library.utm.my:8080/vital/access/manager/Repository/vital:146229 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Quantum ESPRESSO (QE) is a computer simulation package based on Density Functional Theory (DFT) for calculating electronic and structural properties of a material at ground state, which gives an excellent balance of accuracy and computational cost. For a macromolecular system with a large number of atoms, it takes several hours to execute even a simple calculation. The integration of parallel library has made the package compatible to distributes work on many processors through the use of MPI. The computational cost is still challenging as single computer have a limited number of processors. A parallel computing environment of multi-nodes computing system called MPI Cluster is set up on a Linux Operating System to minimize the cost by providing more processors for parallelism. This dissertation investigation evaluates the performance of QE on the multi-node cluster system called MPI-Cluster. We distribute various k-points sampling workload over different MPI processors, to measure the speedup and scalability our multi-note cluster system. The result suggests that the improvement to scaling of speedup over many processors is limited only if the number of k-point to parallelize is greater than the number of processors. We also found the limit of speedup for parallelization of bands calculation is partially independent of the number of bands used and is linearly decreases as the number of MPI processors increased. |
|---|