On 21 October 2013 15:19, Andreas Schäfer <gent...@gmx.de> wrote: > Hi, > > the solution depends on the details of your code. Will all clients > send their progress updates simultaneously? Are you planning for few > or many nodes? > > For few nodes and non-simultaneous updates you could loop on the root > while receiving from MPI_ANY. Clients could send out their updates via > MPI_Isend(). > > If you're expecting many nodes, this 1-n schema will eventually > overwhelm the root node. In that case MPI_Gather() or MPI_Reduce() > will perform better. But those require all nodes to participate. > > Things get complicated if you want non-simultaneous updates from many > nodes... > > HTH > -Andreas
Thanks, currently I run a prototype with 32 mpi processes or so. But I would deploy to a larger set later. ===> root process code: I) mpi thread 1. list all n-tuples 2. split list equally for 32 processes 3. scatter 4. loop to evaluate locally f for my section of space 5. reduce II) UI thread ===> compute mpi process node 3. scatter list of ntuples 4. loop to evaluate locally f for my section of space 5. reduce The loops 4 are not naturally in sync. Would you suggest to modify the loop to do a MPI_ISend after x iterations (for the clients) and MPI_IRecv on the root? Thanks MM
