WAVE_SCHEME = FD

GaussianWavePacket

• 1 CPU core on Forerunner1 = 0.8 CPU core on spock
• 1 node on Forerunner1 = 2.6 nodes on spock = 2.1 nodes on spock with GPU

IsolatedHalo

• 1 CPU core on Forerunner1 = 0.7 CPU core on spock
• 1 node on Forerunner1 = 2.4 nodes on spock = 1.1 nodes on spock with GPU

WAVE_SCHEME = GRAM FE

GaussianWavePacket

• 1 CPU core on Forerunner1 = 0.6 CPU core on spock
• 1 node on Forerunner1 = 2.2 nodes on spock

IsolatedHalo

• 1 CPU core on Forerunner1 = 0.6~0.7 core on spock
• 1 node on Forerunner1 = 2.3 nodes on spock = 1.5 nodes on spock with GPU

Note

• Here only list the time spent on the fluid and gravity solvers. The overall performance may be worse because of the parallel efficiency of other parts (e.g. auxiliary functions).
  • The time is recorded in Record__Timing.
    • See https://github.com/gamer-project/gamer/wiki/Simulation-Logs%3A-Record__Timing for details.
• For WAVE_SCHEME = FD, LAPLACIAN_4TH is on
• For WAVE_SCHEME = GRAM FE, GRAMFE_SCHEME = MATMUL
• For the GaussianWavePacket test problem
  • No AMR is used.
  • N0^3 = 1024^3 → NPatch = 2097152
• For the IsolatedHalo test problem
  • Whole-box Lv1 refinement is used.
  • N0^2 = 512^3
  • N1^3 = 1024^3
• See https://man.twcc.ai/@f1-manual/manual for further information on Forerunner1.
• See https://github.com/calab-ntu/gpu-cluster/wiki/System-Overview#Spock for further information on spock.
• For Forerunner1,
  • We are using the installed gnu compiler.
  • One of the Record__Note
  • We are changing the number of OpenMP threads when the number of CPUs is less than 7.
  • We are changing the number of MPI ranks when the number of CPUs is larger than 7.
  • In some realistic runs of halo evolution, the Flu_Adv + Gra_Adv time becomes 0.14x from using 1 node to 10 nodes.
• For spock,
  • We are using the default intel compiler.
  • One of theRecord__Note
  • We are changing the number of OpenMP threads when the number of CPUs is less than 8.
  • We are changing the number of MPI ranks when the number of CPUs is larger than 8.
• The GPU on spock can make the gravity solver about 3 times faster. However, the overall accelerating effect of GPU depends on the percentage of time that is spent on the gravity solver.
• These tests were done in 2024/08/01~2024/08/14.