Question about MILP parallelization

Hello community,

I have a general question about the branch and bound performance when operated on >1 cores.
I recently ran a rather large optimization problem (100000+ decision variables, 1000+ integers) on 30 threads. What struck me about the computation time was how it developed over the solution process. The root relaxation solved in 16 hours, 3 nodes in 63 hours and 7 nodes in 139 hours. If you break this down, you get a more or less constant calculation time of about 20 hours per node.

Now, I know that this is a very simplified analysis since every node poses a slightly different problem, but still I would expect the calculation time per node to go down with increasing number of nodes due to the parallelization. From my simplified analysis, I would conclude that the nodes are solved sequentially, not parallel.

My question is now if I'm missing something here or if you would draw the same conclusion?

Best wishes
Lukas

Question about MILP parallelization

Comments

Didn't find what you were looking for?