Some questions about Decima's GNN about decima-sim HOT 2 CLOSED

Thanks for your questions.

Re message passing: in our problem, the jobs modeled as directed graphs (child stages cannot run until all parent stages have completed). Passing message from leaves (child stages) to roots (parent stages) have natural meaning - an agent wants to plan ahead and take into account all child stages when deciding to schedule a parent stage. For general pytorch geometric library, my understanding is their generic underlying graph is undirected or bidirectional (e.g., chemical structures, social networks).

Re num_steps/max_depth: I think your understanding is correct. It controls the maximum number of hops a message can traverse in the graph. For example, a single chain graph with 5 nodes A -> B -> C -> D -> E, if the max_depth is 3, then the information encoded in leaf E can traverse up to node B.

Hope these help!

from decima-sim.

Thank you very much for your help.

from decima-sim.