Benchmarking the parallel versions

Firstly, some of the pre-processing was to large for HECToR, even when ran in a small parallel queue, thus most pre-processing was performed on Ness.

The method of decomposition was to set the virtual process topology using np_x=2, np_y=2 and np_z=1,2,4,..., using a simple decompositition method. This ensured that we had a balanced load-balance for the x-y plane at least.

The timing results for the 3 cases are presented graphically in figure 5 and presented in table 5, where can see that the code scales well for all 3 cases, where the optimum number of cores is 64, 256 and 512 for the coarse, medium and fine cases, respectively.

Figure: Time (secs) for the three 3D Jet Break Up cases, where jet_fine, jet_medium and jet_coarse are the times for the fine, medium and coarse cases, and jet_fine_linear, jet_medium_linear and jet_coarse_linear, are their respective perfect scaling curves.

Table 5: Timing and performance results for 3D Jet Break Up

Number	Coarse Mesh	Regular Mesh	Fine Mesh
of cores	Time (Perf)	Time (Perf)	Time (Perf)
4	417.0 (-)	3630.5 (1.18)	- (-)
8	211.1 (1.98)	1804.6 (2.01)	- (-)
16	105.6 (2.00)	900.9 (2.00)	12931.4 (-)
32	57.7 (1.83)	445.4 (2.02)	6568.7 (1.97)
64	27.2 (2.12)	220.5 (2.02)	3354.1 (1.96)
128	20.6 (1.32)	114.7 (1.92)	1752.0 (1.91)
256	24.5 (0.84)	70.9 (1.62)	947.6 (1.85)
512	- (-)	51.6 (1.37)	546.8 (1.73)
1024	- (-)	- (-)	404.9 (1.35)
2048	- (-)	- (-)	894.9 (0.45)