paicli 0.5.4

Client for PAI

paicli

A CLI tool for OpenPAI, which supports basic opperations like listing up jobs, submitting a new job, suspending a running job, and executing ssh into a running container.

How to install

One can install paicli from PyPI.

$ pip install paicli

Basic usage

For more detailed usage, one can use --help option for each subcommand like pai config --help.

$ pai --help
Usage: pai [OPTIONS] COMMAND [ARGS]...

  A CLI tool for OpenPAI.

Options:
  --help  Show this message and exit.

Commands:
  config  Write your configuration to a file.
  host    Show host information of the specified job.
  jobs    Show job list.
  ssh     SSH into a running container.
  stop    Stop a running job.
  submit  Submit your job.
  token   Generate a new access token

Configuration

• Config file: To initialize your config file, run pai config. Then your config file will be located in $HOME/.paicli. You should set your host, port, and username.
• Access token: Before submitting/stopping a job, you should issue your access token by executing pai token and entering your password. You can skip entering password everytime if you write your password directly to your config file (not recommended).

Practical examples

One can utilize and combine the paicli subcommands to achieve several practical operations.

1. Submit multiple jobs with one line

Use pai submit and some template engine like envsubst.

$ head template.json -n 2
{
  "jobName": "template-${JOBID}"
$ for i in `seq 1 3`; do cat template.json | JOBID=$i envsubst | pai submit; done

2. Stop multiple jobs with one line

Combine pai jobs and pai stop.

$ pai jobs -u sotetsuk -s RUNNING | grep template | awk '{print $1}' | xargs pai stop

3. Ssh into multiple running containers and run the same command

Combine pai jobs and pai ssh. In this example, show python processes in multiple jobs with one line.

$ pai jobs -u sotetsuk -s RUNNING | grep template | awk '{print $1}' | xargs -n 1 pai ssh -c "ps -aux | grep python"

4. Show all tensorboard URLs in running jobs

Combine pai jobs and pai host to show all tensorboard URLs. One can use some browser extention to open all URLs (e.g., OpenList)

$ pai jobs -u sotetsuk -s RUNNING | grep template | xargs -n 1 pai host | grep tensorboard | awk '{printf "http://%s:%s\n",$2,$4}'
http://10.0.0.1:9999
http://10.0.0.2:9999
http://10.0.0.3:9999

Author

Sotetsu KOYAMADA