TLaunch: Launch Programs on Multiple Hosts

Introduction

Deepmind launchpad is a library that helps writing distributed program in a simple way. But currently it only supports (or has only open-sourced) launching programs on a single host, either multi-threaded or multi-processed. This library provides a way of launching existing launchpad programs on multiple nodes. Only some simple modification to your program is needed.

Installation

First, let's clone this repository. Then cd into the repository, and execute:

pip install -r requirements.txt
pip install .

Usage

1. Launch programs on multiple hosts without communication

from absl import app
import logging
from tlaunch import lp_ssh

class Worker:
  def __init__(self, worker_id):
    self.worker_id = worker_id

  def run(self):
    logging.info('Worker {}:{}'.format(self.worker_id, i))
    lp_ssh.stop()

def make_program():
  program = lp_ssh.Program('worker')
  worker_num = 2
  current_num = 0
  for host in ['host1','host2']:
    for i in range(worker_num):
      ssh_node  = lp_ssh.SSHNode(Worker, current_num).to_host(host)
      current_num += 1
      program.add_node(ssh_node, label=host+'_worker')
  lp_ssh.launch(program, terminal='ssh_tmux_session')
def main(_):
  make_program()

if __name__ == '__main__':
  app.run(main)

In this code, we place Worker on host1 and host2 via to_host() function. With lp_ssh.launch(), Each Worker will start to run on its corresponding hosts. Besides, examples/mnist/run.sh shows an example of how to train MNIST dataset on multiple hosts.

2. Launch programs on multiple hosts with communication

examples/commands/run_cmd.py gives an example of how to check GPU status of remote hosts. The information can be transferred via defining a TransmitNode.

3. Different data-transfer types

• Star type (see more in examples/trans_types/star_type.py):

  

• Tree type (see more in examples/trans_types/tree_type.py):

  

• Net type (see more in examples/trans_types/net_type.py):

  

[Optional] Kubernetes Support

If you want to use TLaunch with Kubernetes:

1. go installed on host machine to run kustomize.
2. A running kubernetes cluster.
3. Volcano scheduler installed to enable gang scheduling. This will tell you how to install volcano for your kubernetes cluster.

Install and run lp-operator on your kubernetes cluster

cd lp-operator
make deploy

Then lp-operator should be running in namespace lp-operator-system. You can use kubectl get all -n lp-operator-system to check the status of the running operator.

[Optional] Kubernetes Usage

Check here for more information.

[Optional] TPods Usage

1.使用SSH登录TPod

TPod是一款面向分布式场景，为TLaunch准备的用户及资源管理工具。当管理员使用TPod创建用户后，可以自定义的为其指定分配的系统资源（包括CPU、GPU、Memory、Storage），并为该用户创建一个已经预装好TLaunch框架的TPod开发机，用户可以直接通过SSH登录该机器访问集群，以快速进入开发流程。

TPod的存储结构

在TPod中，我们会在/TData目录下创建你的个人文件夹，该文件夹会通过挂载的文件系统与远端同步。/TData内会预先创建以下内容：

• code:用于存放训练代码
  • setup.sh:用于指定训练环境及代码的安装方法
• data:用于存放训练所需的数据
• cache:用于存放分布式计算过程中产生的缓存文件
• models:用于存放模型数据

  2.存放算法代码并指定安装方法

  在多数分布式场景中，少量的代码往往会经常改动。若每次改动代码都重新构建镜像，会浪费大量的时间。因此，我们可以将这部分代码存放在/TData/code文件夹中。在训练中，每当一个pods被创建时，都将先按照code文件夹中的setup.sh脚本更新环境

  3.调用TLaunch创建任务

  Check here for more information.

  4.管理任务

  当任务创建完成后，我们可以调用kubectl来查看任务状态及日志，其中常用的几条指令包括：

• 查看正在运行中的任务：kubectl get lpjobs
• 查看正在运行中的pods:kubectl get pods
• 查看pod节点日志:kubectl logs ${pod name}
• 查看pod节点详细信息:kubectl describe pods ${pod name}
• 删除任务:kubectl delete lpjobs ${lpjob name}

Citing TLaunch

If you use TLaunch in your work, please cite us:

@article{tartrl2021tlaunch,
    title={TLaunch: Launch Programs on Multiple Hosts},
    author={Shiyu Huang, Zeming Liu, Shouren Yang, Shizhen Xu, Ting Chen, Jun Zhu, Sen Na},
    year={2021},
    howpublished={\url{https://github.com/TARTRL/TLaunch}},
}