search

eugenevinitsky / sequential_social_dilemma_games

Repo for reproduction of sequential social dilemmas
MIT License
384 stars 134 forks source link

Increasing the number of nodes does not speed up train_baseline #158

Open gregSchwartz18 opened 5 years ago

gregSchwartz18 commented 5 years ago

I am trying to run train_baseline on a cluster. When I increase the number of CPUs used per node the experiment speeds up. But when I use more nodes, while keeping the CPUs per node constant, the speed stays the same (disregarding the longer initial iteration). Is this repo set up to use multiple nodes? If not, what is the best way to go about adding that functionality?

Looking at the output logs, I also noticed that only iterations 1, 4, 7, and 10 were being printed. Why might this be happening?

train_baseline on 1 node.txt train_baseline on 2 nodes.txt train_baseline on 5 nodes.txt

eugenevinitsky commented 5 years ago

Sorry about the delay! This repo is set up to use multiple nodes: are you increasing the number of workers? It won't use more cpus than num_workers. As for the logging, I'm not sure, you may want to post on the rllib github.