GATE simulations on multicomputer architectures

GATE/Geant4 simulations are computationally demanding, requiring thousands of processor hours to produce realistic results. One of the difficulties in producing a truly parallel solution is that in PET simulations, events are NOT independent of each other completely. Photons that originate in a given time window must be simulated in the same processor so that they can be fed to the same coincidence processing chain. Otherwise, a centralized coincidence processing approach would incur in excessive communication overhead. The cluster solution for GATE was a first solution to tackle this problem, splitting the simulation in very large windows. Here we propose a truly parallel computational model that can handle singles and coincidences in a natural way, producing simulation results equivalent to its sequential counterpart. We show that the MPI implementation of this model speeds up almost linearly with the number of processors.