Test Case 1
The system simulated is a hydrogen defect in 32 atoms of palladium. It uses 10 k-points and the PBE exchange correlation functional ([27]-[28]).
Table:
Scaling of Test Case 1.
Test Case 1 |
Cores |
Time (secs) |
Speedup |
VASP 5.2.2 |
64 |
298.187 |
1 |
KPAR=2 |
128 |
159.982 |
1.863 |
KPAR=5 |
320 |
75.357 |
3.956 |
KPAR=10 |
640 |
47.795 |
6.239 |
|
Table (
) shows that the k-point parallelized code scales satisfactorily to 320 cores, where it is twice as fast as the original code at the same number of cores (see Table (
)).
Table:
The optimal NPAR for the original code is 4, 8, 32, and 32 for 64, 128, 320 and 640 cores respectively. For the k-point parallelized code the optimal NPAR is 4 for any number of cores, provided one k-group consists of 64 cores.
Test Case 1 |
Cores |
Time (secs) |
Speedup |
VASP 5.2.2 |
128 |
206.517 |
1 |
KPAR=2 |
128 |
158.686 |
1.301 |
VASP 5.2.2 |
320 |
146.197 |
1 |
KPAR=5 |
320 |
75.201 |
1.944 |
VASP 5.2.2 |
640 |
147.606 |
1 |
KPAR=10 |
640 |
47.795 |
3.088 |
|
Figure 1:
Speedup for Test Case 1 (where Speedup is taken to be 1 for 64 cores).
![\includegraphics[angle=0,width=14cm]{TC_1.eps}](img6.png) |
Asimina Maniopoulou 2011-07-09