Name: CoMD-CUDA
Owner: NVIDIA Corporation
Description: GPU implementation of classical molecular dynamics proxy application.
Created: 2013-11-07 22:26:06.0
Updated: 2018-03-29 10:19:41.0
Pushed: 2017-01-30 15:37:07.0
Size: 9706
Language: C
GitHub Committers
| User | Most Recent Commit | # Commits |
|---|
Other Committers
| User | Most Recent Commit | # Commits |
|---|
This is GPU implementation of CoMD 1.1 proxy application. The GPU code supports both cell-based and neighbor-lists methods for molecular dynamics and includes various parallelization strategies for both. Distributed multi-GPU implementation is supported as well by using one GPU per MPI rank.
Use src-mpi/Makefile.EAM for EAM forces and src-mpi/Makefile.NAMD for LJ forces. Modify Makefiles according to your environment, then make.
Requirements:
Notes:
Single-GPU run with EAM forces and 49x49x49 grid size using default method (cell-based, thread per atom):
bin/CoMD-cuda-mpi -e -x 49 -y 49 -z 49
Multi-GPU run with 2 GPUs, EAM forces and 98x49x49 overall grid size divided between GPUs along X dimension:
irun -np 2 ./bin/CoMD-cuda-mpi -e -x 98 -y 49 -z 49 -i 2
Multi-GPU run with 2 GPUs, EAM forces, 98x49x49 overall grid size and neighbor lists method:
irun -np 2 ./bin/CoMD-cuda-mpi -e -x 98 -y 49 -z 49 -i 2 -m thread_atom_nl
Best single-GPU configuration using warp per atom approach with neighbor lists:
bin/CoMD-cuda-mpi -e -x 49 -y 49 -z 49 -m warp_atom_nl
To view all available options please check:
Below is a sample output which you can use for the validation of the results. When modifying the code please check that all energies and # of atoms remain the same as in the original code.
Performance
oop Time(fs) Total Energy Potential Energy Kinetic Energy Temperature (us/atom) # Atoms
0 0.00 -3.460523233086 -3.538079224686 0.077555991600 600.0000 0.0000 470596
10 10.00 -3.460522622766 -3.529929454580 0.069406831814 536.9553 0.0707 470596
20 20.00 -3.460524220490 -3.509740515517 0.049216295027 380.7543 0.0711 470596
30 30.00 -3.460527806915 -3.488529040692 0.028001233777 216.6272 0.0660 470596
40 40.00 -3.460532196608 -3.477523402265 0.016991205657 131.4498 0.0662 470596
50 50.00 -3.460536497383 -3.479780609997 0.019244112614 148.8791 0.0709 470596
60 60.00 -3.460538213894 -3.488976046432 0.028437832538 220.0049 0.0665 470596
70 70.00 -3.460536800219 -3.496688002423 0.036151202204 279.6782 0.0663 470596
80 80.00 -3.460533977439 -3.498984084647 0.038450107208 297.4633 0.0713 470596
90 90.00 -3.460531463100 -3.497356126200 0.036824663100 284.8883 0.0664 470596
100 100.00 -3.460530040624 -3.495833910540 0.035303869916 273.1230 0.0666 470596
Performance metric for CoMD is atoms/us (processed atoms per time), which is printed at the end of the execution.
-----------------------------------------------
rage atom rate: 14.66 atoms/us
-----------------------------------------------
EAM code can achieve 34 atoms/us on NVIDIA Tesla K40m with the boost clocks.
GTC 2014: Optimizing CoMD: A Molecular Dynamics Proxy Application Study
CoMD is a reference implementation of typical classical molecular dynamics algorithms and workloads. It is created and maintained by ExMatEx: Exascale Co-Design Center for Materials in Extreme Environments (exmatex.org). The code is intended to serve as a vehicle for co-design by allowing others to extend and/or reimplement it as needed to test performance of new architectures, programming models, etc.
Original CoMD code is available at github.com/exmatex/CoMD.
To view the generated Doxygen documentation for CoMD, please visit exmatex.github.io/CoMD/doxygen-mpi/index.html.
To contact the developers of CoMD send email to exmatex-comd@llnl.gov.