The Ohio State University   |   Department of Physics

OHMMS  (Object-oriented High-performance solutions for Multi-scale Materials Simulations)

The core components of OHMMS are based on PETE and other numerical and parallel libraries.  The most common applications of OHMMS are atomistic simulations of crystalline and molecular systems using various model Hamiltonians to describe the interaction of constituent particles.  OHHMS and its related softwares are the products of research efforts by Jeongnim Kim and her collaborators at National Center for Supercomputing Applications and Materials Computation Center at Universtity of Illinios, Urbana-Champaign and the Department of Physics, The Ohio State Universtity.

OHMMS achieved i) high flexibility and re-usability by applying object-oriented programming practice, such as design patterns and ii) performance by utilizing advanced abstractions like expression templates for low-level components.  In particular, applications of OHMMS can utilize MPI, OpenMP or MPI/OpenMP parallelizations to meet their special needs depending on the algorithms and comupter architectures.

 

OHMMS vs LAMMPS (serial performance)

Tests were performed on the Glenn cluster at OSC (single processor : AMD Opteron 2.6 GHz).  During tests, no output was produced.

Number of Si

Time for 10000 MD steps

Memory usage

min

%

Mb

OHMMS

LAMMPS

speed-diff

OHMMS

LAMMPS

1000        

0.9          

0.8          

9.3          

14          

1.7         

8000        

7.6          

5.6          

26.0          

58          

5.6         

64000        

64.9          

46.7          

28.1          

410          

36            

512000        

479.4          

384.3          

19.8          

3234          

261            

1000000        

-

726.3          

-

-

502            

  • Both show the linear behavior in their speed and memory usage.
  • LAMMPS runs about 20 % faster than OHMMS, and uses only 10 % of the memory of OHMMS.

 

Parallel efficiency of OHMMS (OpenMP)

The parallel efficienct of the current version of OHMMS is about 0.56.

Number of Si

Time for 40000 MD steps in min

Serial

OpenMP- 4 processors

OpenMP- 8 processors

Total time

Wall-time

Efficiency

Total time

Wall-time

Efficiency

1000        

3.5   

4.3    

1        

0.8          

6.3    

0.8     

0.56

8000        

30     

47       

12        

0.64        

58       

7        

0.52

64000        

242     

445       

111        

0.54        

416       

52        

0.58

512000        

1917     

2942       

736        

0.65        

3435       

429        

0.56