Optimal combination of number of environments and number of threads

[shafeef901]

What would be the optimal combination of the number of environments and the number of threads while using raisimGymTorch? My computer has 20 cores and I've tried num_envs=20 and num_threads=20. However, it's slower than using higher num_threads (30/40). Do you have any suggestions on this? FYI, I am using RaiSim for RL using torch and gym.

[jhwangbo]

In the end, you have to tune them for the best results. Bigger the num_envs, the higher FPS you will get. But too big num_envs will eventually slow down the training in terms of performance vs training time. I also found the optimal num_threads to be double the number of cores as well. This is proably due to hyperthreading but I am not sure.

[ClaudioChiariello]

Excuse me, I have a trivial question. If I have 8 core, , each capable of running 2 threads, I should set num_thread = 16 to maximize performance?

And this should mean that I can parallelize up to 16 environments. However, if I set num_envs=50, since it can only parallelize a maximum of 16 threads, it will launch the first 16 environments first, then the next 16, then the next, and finally the last 2. Is that correct?

I mean, it serialized the environment that cannot parallelize. I'm not entirely certain if I'm stating things correctly

[jhwangbo]

"I should set num_thread = 16 to maximize performance?": This is often true but you have to tweak the number +-10% to get the best performance "And this should mean that I can parallelize up to 16 environments. However, if I set num_envs=50, since it can only parallelize a maximum of 16 threads, it will launch the first 16 environments first, then the next 16, then the next, and finally the last 2. Is that correct?": This is almost true. The only incorrect part is that the threads do not wait for each other. They don't work in a batch. If one thread finishes first, it will automatically grab the next job.

[ClaudioChiariello]

Thank you for the fast reply