[gate-users] Re: gate-users Digest, Vol 20, Issue 1

[patrice descourt]

Dear Vandana

you do not have to bother about installing geant or gate on each node, 
MPI logs on each node via rsh or ssh depending on your parameterization.

Below are outlines along which you can get parallel jobs


I will be pleased to give the final codes in a directory when I have 
written those because for now I only focus on the second implementation 
which is far more cumbersome.

The first one should not present any problems :

You just have to install MPI and compile Geant and Gate with mpicc which 
replaces gcc

and run Gate via mpirun :

mpirun -np #number of nodes  your_compiled_code.exe

you have to include

#include "mpi.h"  in each file using MPI library e.g Gate.cc and 
G4RunManager.hh




In Gate.cc you need at the beginning of main()

to add

MPI::Init(argc,argv);   : this is to initialize MPI

at the end of Gate.cc main()

you put

MPI::Finalize();


in G4RunManager::DoEventLoop()

you put

G4int proc_number = MPI::COMM_WORLD.Get_size();
G4int proc_ID = MPI::COMM_WORLD.Get_rank();

before the for loop

 which you replace with

for(i_event = proc_ID; i_event<n_event; i_event += proc_number)

now for the output,  I review the ROOT one but the scheme is the same 
for any GateVOutputModule :

in GateToRoot.hh you put #include "mpi.h"

and put at the beginning of GateToRoot:RecordBeginOfAcquisition

G4int proc_ID = MPI::COMM_WORLD.Get_rank();

and replace each GetFilePath() by  GetFilePath()+"proc_"+proc_ID

I will give all the codes when this is all implemented

Cheers
P.Descourt