Hybrid-View Programming of Nuclear Fusion Simulation Code in XcalableMP
193
Table 4 Process mapping to
evaluate the scaling of
decomposition for each
domain (N t xN r xN rp )
Processes Toroidal
Radial
Particle
16
2 × 2 × 4 2 × 4 × 2
2 × 2 × 4
32
4 × 2 × 4 2 × 8 × 2
2 × 2 × 8
64
8 × 2 × 4 2 × 16 × 2
2 × 2 × 16
128
16 × 2 × 4 2 × 32 × 2
2 × 2 × 32
256
32 × 2 × 4 2 × 64 × 2
2 × 2 × 64
512
64 × 2 × 4 2 × 128 × 2 2 × 2 × 128
radial and particle dimensions as 2 × 4. This was due to variations in the number
of calculations because increasing the toroidal dimension also changes the poloidal
planes, as described above. We used this scheme to change the scaling dimension.
Second, we used 16 nodes where one process ran on each node and the number of
threads ranged from 1 to 16 in each process. The processes mapped on each domain
to evaluate the decomposition are 2 × 4 × 2 and 2 × 2 × 4.
4.2 Results
With weak scaling, Figs. 12, 13, and 14 shows the elapsed time for both calculation
and communication of MPI, XMP-localview, and XMP-hybridview required to
scale the number of processes from 16 to 512, where decomposition on the toroidal
Fig. 12 Elapsed time of the decomposition on toroidal dimension from 16 to 512 processes in
weak scaling