Andrea: From the documentation I found the flow for the creation of an application code is the following: 1. Create a Network and register components 2. Set network size and layout 3. Set components and number of variables for nodes/edges 4. Set up the network 5. Distribute the network
Following that flow, from point 1 to point 4 the Network is "owned" only by the 0 rank processor while after point 5 the network is partitioned and "owned" by the n processors involved in the computation (according to the partitioner used). Correct. Currently, DMNetwork only supports sequential setup of network. I just started parallelizing it. I'll let you know when I'm done. I was trying to let all the n processors involved read a portion of the network in such a way that the flow listed before will be a multiprocessor process and not restricted only to the rank 0 processor (parallel input data). In this way after point 4 I will have a distributed network and I don't need to call DMNetworkDistribute(). Unfortunately the documentation I found doesn't explain how to do that, so I would like to ask you if it is feasible and if you have some examples or documentation that you can share for this purpose. It is not supported yet. Internal works are needed. You may follow my development branch https://bitbucket.org/petsc/petsc/branch/hzhang/pdmnetwork-setup If you have idea/suggestions, let us know. Hong
