Changes between Version 23 and Version 24 of Public/User_Guide/Offloading_hybrid_apps


Ignore:
Timestamp:
Sep 18, 2019, 9:27:00 AM (5 years ago)
Author:
Kevin Sala
Comment:

Legend:

Unmodified
Added
Removed
Modified
  • Public/User_Guide/Offloading_hybrid_apps

    v23 v24  
    9898group. Thus, each MPI process is in charge of computing the forces or updating the particles
    9999of a consecutive chunk of blocks.
     100
     101This computation pattern is repeated during multiple timesteps. The communication pattern
     102in each timestep consists in the fact that GPU processes exchange their particles with each
     103other in a circular manner to compute the forces between their particles against the particles
     104of all other GPU processes. For simplifying this pattern, the application uses a different MPI
     105communicator for this circular exchanging. Once a GPU process finishes the computation of its
     106forces, it sends the forces to the corresponding CPU processes, and then, it receives the
     107updated particles. We perform separate sends/receives for each block.
    100108
    101109The available versions are: