cudaMPI is Message Passing for GPGPU clusters.  See http://www.cs.uaf.edu/sw/cudaMPI/ .  The author is Orion Lawlor.

This combines the MessagePassingInterface (MPI) with the ComputeUnifiedDeviceArchitecture (CUDA) to provide a means where more than one GPGPU can share in a computation task.

The web site has a link to a technical paper with information on the timing of data transfers to and from GeneralPurposeGraphicsProcessUnits (GPGPUs).

----
ParallelProgrammingModel CategoryCee CategoryCpp CategoryGpgpu