Test Case 3

The simulated system is PbO using 126 k-points. The original code exhibits slowdown for over 144 cores and the only way to employ more cores is to use the k-point parallelized code. The latter scales quite satisfactorily up to 1008 cores.

Table: Scaling of Test Case 3.

Test Case 3	Cores	Time (secs)	Speedup
VASP 5.2.2	144	120.388	1
KPAR=2	288	69.208	1.740
KPAR=3	432	48.315	2.492
KPAR=6	864	30.234	3.982
KPAR=7	1008	27.402	4.393
KPAR=9	1296	40.443	2.976

Table: Here the optimal NPAR for the original code is 18, 12,18, 36 for 144, 288, 432, 1008 cores respectively. For the k-point parallelized the optimal NPAR for any number of cores is 18, provided a k-group consists of 144 cores.

Test Case 3	Cores	Time (secs)	Speedup
VASP 5.2.2	288	130.688	1
KPAR=2	288	69.208	1.888
VASP 5.2.2	432	122.236	1
KPAR=3	432	48.315	2.530
VASP 5.2.2	864	320.196	1
KPAR=3	864	30.234	10.591
VASP 5.2.2	1008	318.516	1
KPAR=7	1008	27.402	11.624

Figure 3: Speedup for Test Case 3 (where Speedup is taken to be 1 for 144 cores.)