CPU resources, htop shows all 12 cores being used for example_1.py

Thanks for the great library. I am running some tests to benchmark performance. Since I hope to be running these using many CPUs, I want to understand how many CPUs a script will consume. I installed the repository as of today, and ran:

python examples/example_1.py

I’m running this on a machine with a single GPU, and an i7-8700k CPU with 12 logical cores. I assume I’m not using the GPU in the above command.

In a separate tab, my htop is regularly showing something like this:

All 12 CPUs on my machine are being used.

When I run:

python examples/example_1.py --cuda_idx 0

I don’t generally see all 12 CPUs being used, but I may get something like 6 CPUs used:

Just wondering, the documentation of example_1.py says this:

Runs one instance of the Atari environment and optimizes using DQN algorithm.
Can use a GPU for the agent (applies to both sample and train). No parallelism
employed, so everything happens in one python process; can be easier to debug.

However I assumed “one python process = one core”. Perhaps this is not the right way to think about it. Is there a way to roughly estimate how many CPUs (or “cores” – I use terms interchangeably) will be used for a given training run?

[DanielTakeshi]

Thanks @astooke

I did another test where I took example_5 and did one change, which is to replace MinibatchRlEval with MinibatchRl so that there’s no evaluation environment. I then ran it with 8 parallel envs as you can see below. htop is regularly showing at least 24 CPUs being used (on a 48 CPU machine) and nothing but my script is running on it.

Let me now do some testing with your suggestion of changing the num threads …

Update 1: was going to do some experimentation but saw @codelast put some new stuff below.

Update 2: actually I might have gotten confused. In example_5.py the batch_B is hard-coded at 32 so the n_parallel does not actually change it. I assumed n_parallel was the number of parallel environments, but I guess it’s more viewed as the amount of resources provided to a given program? I will continue investigating but do you have any quick suggestions / comments on the distinction between n_parallel and batch_B?

Update 3: figured out my prior question. batch_B is actually the number of parallel environments, and n_parallel is the number of workers that need to run those environments. If batch_B=32 and n_parallel=2 then we have [16,16] envs allocated to the two workers. If n_parallel=3 then it’s [11,11,10] envs allocated to the three workers, and we get a warning saying that performance may suffer due to unequal environment distribution.

[DanielTakeshi]

I just ran some tests today, and indeed using taskset helped to limit my CPU usage, according to htop.