Running on a 6 core machine with 12 logical processors, and I can’t really figure out an optimal balance between num-envs and num-areas for the best training result. My thinking is maybe to set num-envs == number of cores and then tweaking the num-areas such that the steps finish the quickest but wondering what approach everyone else is using?
I think it depends on the unity environment + training configuration.
For my current project I first maxed out the singe core that was doing most of the work the single (unity) environment by introducing more areas, which in my case was 4. Then I started benchmarking the fastest training by trying out the optimal number of environments in this case it was 6.
I ended up with --num-areas 24 --num-envs 6
I think it’s right… seemed to have the best training speed / top cpu utilization, but would love to have someone from Unity chime in as it’s not all that well documented as of yet.