Let's say we have a cluster of 20-2000 heterogenous compute nodes. Consider for example the parallel solution of the helmholtz equation: Now we want to distribute the solution process and, to make things easier, we split the problem in a fine-grained way (partial solution of the system matrix). We could train an Ai with the time taken to solve the subproblem depending on multiple factors (for example, size of the mesh, needed precision, etc) and let the Ai choose the optimal distribution and division of the problem based on the available data.
I'm new to the area of Artificial Intelligence. Are there any open source frameworks which could accomplish this task? How would you estimate the required amount of compute power to train the network?