-->

These codes illustrate how to use a hybrid shared/distributed memory algorithm, with a tiled scheme on each shared memory multi-core node implemented with OpenMP, and domain decomposition connecting such nodes implemented with MPI. The algorithms are described in detail in Refs. [2-4].

For the 2D electrostatic code, a typical execution time for the particle part of this code is about 80 ps/particle/time-step with 576 processing cores. For the 2-1/2D electromagnetic code, a typical execution time for the particle part of this code is about 230 ps/particle/time-step with 576 processing cores. For the 2-1/2D Darwin code, a typical execution time for the particle part of this code is about 364 ps/particle/time-step with 576 processing cores. The CPUs (2.67GHz Intel Nehalem processors) were throttled down to 1.6 GHz for these benchmarks.

For the 3D electrostatic code, a typical execution time for the particle part of this code is about 110 ps/particle/time-step with 768 processing cores. For the 3D electromagnetic code, a typical execution time f or the particle part of this code is about 280 ps/particle/time-step with 768 processing cores. For the 3D Darwin code, a typical execution time for the particle part of this code is about 715 ps/particle/time-step with 768 processing cores. The CPUs (2.67GHz Intel Nehalem processors) were throttled down to 1.6 GHz for these benchmarks.

Electrostatic:

  • 1. 2D Parallel Electrostatic Spectral code: mppic2
  • 2. 3D Parallel Electrostatic Spectral code: mppic3

Electromagnetic:

  • 3. 2-1/2D Parallel Electromagnetic Spectral code: mpbpic2
  • 4. 3D Parallel Electromagnetic Spectral code: mpbpic3

Darwin:

  • 5. 2-1/2D Parallel Darwin Spectral code: mpdpic2
  • 6. 3D Parallel Darwin Spectral code: mpdpic3

Figure below: Performance as a function of the number of cores. Dashed red line is ideal scaling, blue shows particle time, black shows total time. Degradation of total time is due to the all to all transpose in the FFT, which for large number of cores is beginning to be dominated by message latency. The degradation is considerably less than in the MPI-only case.

Figure details (click to enlarge): Figure 1; Figure 2; Figure 3; Figure 4.

Want to contact the developer? Send mail to Viktor Decyk at decyk@physics.ucla.edu.

Vectorization Codes