17:15 〜 18:45
[MGI29-P04] スーパーコンピュータ「富岳」にて計算する火成活動―マントル対流結合系の数値シミュレーション
キーワード：富岳、数値解析、マントル対流
To understand the evolution of rocky planetary interiors, it is necessary to consider not only mantle convection but also the thermal and compositional changes caused by magma generation and migration (hereafter referred to as magmatism). The volcanic activity and the global expansion/contraction of planet are strongly influenced by magmatism caused by partially molten plumes. We evolved the efficiency of numerical calculation of mantle convection and magmatism in a 3D spherical shell on both the supercomputer “Fugaku” and a Linux server (Precision 3660, Core i9 with 64GB). A finite difference numerical code calculates the energy, mass, and momentum equations for mantle magmatism and mantle convection in a 3-D sphere [r, θ, φ | 385 km ≦ r ≦ 1735 km, 0 ≦ θ≦ 0.5π, 0.25π ≦ φ ≦ 0.75π ] on a mesh with 64 (radial direction) times 64 (latitudinal direction) times 128 (longitudinal direction) mesh points under the Boussinesq approximation. The inner and the outer radii correspond to the core and planetary radii, respectively. We decomposed the flow field of solid matrix into the solenoidal (divergence-free) and the potential components both of which are solved by the multigrid method; the smoothing calculations of solenoidal components are done by the pseudo-compressibility method, while those of the potential components are by the Gauss-Seidel smoothing together with the multi-color ordering.
The computational time for 1 step on Fugaku is 4.75 sec. while that on the Linux server is 8.40 sec. after the initial transient stage of the calculation. By comparing Fugaku with the Linux server, the overall calculations on Fugaku are only 1.8 times faster than those on the Linux server despite the fact that the theoretical peak performance on 1 CPU of Fugaku (3072 GFLOPS) is around 4.7 times larger than that on the Linux server (902 GFLOPS). Perhaps, the potential of Fugaku may not be brought out owing to the connection time among CPU parallelism in our cases which need extensive communications. As for the calculation of the solenoidal part, on the other hand, the Linux server is 3.8 times faster than Fugaku, which suggests that memory access with a stride larger than one should be avoided on Fugaku. These results help to develop a planetary model for the future.
The computational time for 1 step on Fugaku is 4.75 sec. while that on the Linux server is 8.40 sec. after the initial transient stage of the calculation. By comparing Fugaku with the Linux server, the overall calculations on Fugaku are only 1.8 times faster than those on the Linux server despite the fact that the theoretical peak performance on 1 CPU of Fugaku (3072 GFLOPS) is around 4.7 times larger than that on the Linux server (902 GFLOPS). Perhaps, the potential of Fugaku may not be brought out owing to the connection time among CPU parallelism in our cases which need extensive communications. As for the calculation of the solenoidal part, on the other hand, the Linux server is 3.8 times faster than Fugaku, which suggests that memory access with a stride larger than one should be avoided on Fugaku. These results help to develop a planetary model for the future.