google-deepmind / google-deepmind/dm_control

How can I control the number of CPU threads used by the environment?

未关闭
#329 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看
主要语言
Python
星标
4.7k
派生
764
PR 合并指标
30 天内没有已合并 PR

描述

I notice that some environments in the DMControl suite may take more than one CPU thread.
For example, if we choose Domain_name=finger, Task_name=spin.

I run this task in a server with 40 threads Intel(R) Xeon(R) CPU E5-2640 v4 @ 2.40GHz and 4 Titan XP. This task takes 8 threads when starting. Then it will take 3 CPU threads when train my agent.

I run this task in a server with 32 threads Intel(R) Xeon(R) Gold 6137 CPU @ 3.90GHz and 4Titan RTX. This takes 12 threads when starting. Then it will take 12 CPUs threads as training.

And different tasks may take different threads. For example, if we start Domain_name=ball_in_cup, Task_name=catch, then it takes 16 threads as training.

It is a heavy computational burden if the task takes too much CPU resources.

So, how can I control the number of CPU threads used by the environment of DMControl suite?

Thank you and best wishes.

贡献指南

打开贡献指南

调研方向

Issue 提到了 DMControl 环境 finger/spin 和 ball_in_cup/catch,但没有指出任何文件、测试或入口点。首先复现报告中的 CPU 使用情况,并跟踪环境初始化和训练行为。完成标准是记录一种经过验证的、用于控制或限制环境 CPU 线程的方法。

由索引模型根据 Issue 内容生成。

评估

技术栈
python
领域
machine-learning
Issue 类型
文档
难度
3/5
预计耗时
1-2 天
活跃度
停滞
描述清晰度
需要澄清
新手友好度
25/100

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。