Parallel computing for Reinforcement Learning training on VM

Hi,

I am writing to ask if there's a way to increase the number of vCPU assigned to a worker when using parallel training for Reinforcement Learning application?

I noticed that the number of vCPU used at 100% is the same as the number of workers (set using parpool(numworkers)). When testing my model and running simulations on my local computer, the computational load exceeds the processing power of 1 vCPU. It took approximately 3-4 cores (50% of a typical intel i7 CPU) to run the simulation and train the agent.

Therefore, I would like to increase the number of vCPU assigned to a worker. I've tried to set 'numthreads' to 4 per worker, but that doesn't seem to solve the problem.

I am using a Ubuntu 18.04 Virtual Machine to run Matlab 2021a.

Thank you!

[Emmanouil Tzorakoleftherakis]

Any reason why you do not increase the number of workers instead?

[Tech Logg Ding]

Hi Emmanouil,

Thank you for your reply. I've initially wanted to increase the number of cores per worker as the simulation was quite computationally expensive. However, after learning more about parallel computing, I understand that using more cores per worker might not help the problem if the code is not configured to use more cores. I have increased the number of workers to use more of the available cores at a time and that seems to work fine. I also checked and noticed that the 'client' is configured to use more cores and is able to distribute its' workload across multiple cores.