Skip to content
fomics edited this page Jul 13, 2015 · 7 revisions

PI: David Daverio

Team members:

  • Julian Adamek
  • David Daverio
  • Martin Kunz

Mentors:

  • Michel Mueller
  • Matthew Colgrove

Institutions: University of Geneva, University of Sussex, University of the Basque Country

Application Area: Latfield2, a light-weight framework for classical field theory simulations that transparently and efficiently handles parallelisation on small to large (100k+ cores) computers. The framework includes a fully parallelized FFT and an I/O server. Current 'user codes' of the framework include cosmic string simulations in the Lattice Abelian Higgs model (LAH) and general-relativitistic cosmological N-body simulations (gevolution).

Starting point: Production: 4096^ sites with 16K MPI processes, 13K FFTs, 63% of execution time. Small test problems used here. Goal: to port FFT to GPU and attain at least parity with CPU performance to free up latter for other tasks.

Porting Steps: Ported FFT wrapper based on FFTW and direct calls to cuFFT, manual data management with native cuFFT interface, plus pinded memory and complete OpenACC integeration incl. transposes.

Results: from 8.00s (CPU) to 5.39s (GPU) with. Possibly more

Future: Next project: N-body, must port solvers themselves to GPUs.

Issues: Poor documentation of cuFFT, lost 10 hours.

Final Presentation