hmm this is not right, the decission how many nodes work together in the distributed kernel should be made elsewhere. Perhaps a DistributedAccess instance should be given to a DataFactory Definition at line 258 of file Factory.h. References Impala::Core::Training::LoadDistributedAccess(), mDataSet, mDistributedAccess, mFeature, mModelType, and Impala::Core::Matrix::DistributedAccess::Subscribe(). Referenced by MakeDataSrc(), and ServeDistributedAccess(). 00259 { 00260 if(mDistributedAccess == 0) 00261 { 00262 if(Link::Mpi::NrProcs() == 1) 00263 { 00264 mDistributedAccess = Training::LoadDistributedAccess 00265 (mModelType, mFeature, mDataSet, 0, 0, 1); 00266 } 00267 else 00268 { 00269 mDistributedAccess = Training::LoadDistributedAccess 00270 (mModelType, mFeature, mDataSet, 0, 1, 00271 Link::Mpi::NrProcs() - 1); 00272 if(Link::Mpi::MyId() == 0) 00273 mDistributedAccess->Subscribe(); 00274 } 00275 } 00276 return mDistributedAccess; 00277 }
Here is the call graph for this function:
|