Question about violation of computation time constraint

[JinZQ56]

Hi,

We have a question about the computation time constraint.
We submit our agent to the cloud server. The result shows that our agent violated the computation time constraint many times(400 for hit-task and 700 for defend-task.)
But when we test our agent locally, this is not the case(the number of computation time violations is under 100 for both tasks.)
We only use the CPU to evaluate the agent. The CPU is Intel Core i7-7700HQ.
May I ask what could be the possible reason for the huge difference?
If it is due to the different hardware, what type of CPU is used in the cloud server?

Another question, when I built the docker image according to the documentation of the challenge, I just ignored installing the NVIDIA Container Toolkit. But I can still build the image, test the evaluation locally and submit the agent to the cloud.
Is this toolkit important for the evaluation? Could this be the reason for the increased computation time?

[PuzeLiu]

Hi,
The CPU of the server is: Intel(R) Xeon(R) Gold 6278C CPU @ 2.60GHz.
We use 12 cores for 3 parallel evaluations. So on average, 4 cores will be assigned to each evaluation.
Apparently, the server will take a bit more time than on your local machine. I've also checked your dataset. Apparently, some computation time is still high:

Here is the code, I use to check the time:

import os.path

import numpy as np
import matplotlib.pyplot as plt
from mushroom_rl.utils.dataset import compute_episodes_length

def main():
    data_dir = ""
    env = "3dof-hit"
    dataset = np.load(os.path.join(data_dir, env, "dataset.pkl"), allow_pickle=True)
    computation_time = np.load(os.path.join(data_dir, env, "computation_time.npy"), allow_pickle=True)

    episode_length = compute_episodes_length(dataset)

    current_idx = 0

    max_computation_time = list()
    for episode_len in episode_length:
        max_computation_time.append(np.max(computation_time[current_idx: current_idx+episode_len]))
        current_idx += episode_len

    plt.hist(max_computation_time)
    plt.show()

if __name__ == '__main__':
    main()

One option you could consider is the use of a separate thread for computing the trajectory.