Cloud Berg

Design goal

Berg is a minimal tool for running experiments with GPU instances on Cloud and storing the results in a bucket. It’s 400 lines of python.

Non-goals

not designed to run on anything beside google cloud
not designed to host a production web server
not designed to do anything clever

Installation

Clone and install the berg tool on your local machine After installing gcloud, run the following locally

gcloud init
gcloud auth login # Use your @google account
gcloud source repos clone berg --project=cloud-berg
pip install -e berg
berg init # Set up your configuration correctly in ~/.berg/berg.json

Basic usage

Launch two experiments (that each use all the GPUs on single instance, and shut down after they complete their job)

cd <your_git_repo>
berg run "python train.py --bs=256 --logdir=/root/berg_results/sweep1/bs_256"

The python command will be run from the root of your github repo. It will run rsync files from /root/berg_results/sweep1/ and results will show up in gs://<your_berg_bucket>/berg_results/sweep1

Launch a devbox with 4 p100 gpus, CUDA, Tensorflow and Pytorch

berg devbox

After it starts you can ssh into it with berg ssh <instance_name>

How does berg work?

The best way to understand berg is to just read the code (it's very small). Here’s an even shorter TLDR:

Executing berg run does the following:

Copy the local git repo that you’re working on to GCS
Generate a job-specific <job_name>_metadata.json file with the cmd and the current git_commit and copy it to GCS
Start an instance to process that job (using the config info in ~/.berg/berg.json)

Once the instance starts, it executes berg-worker run which does the following:

Copy down <job_name>_metadata.json from GCS and parse the file
Copy down the git repo from GCS, check out your commit and install any requirements in setup.py
Start regularly uploading logs from /root/berg_results/<results_dir>
Run the cmd for the job
Shut down the instance after the job finishes

Managing berg jobs

Helpful shortcuts (all of which are aliases for gcloud commands)

berg list # list all running instances 
berg tail <inst_name> # tail logs
berg log <inst_name> # get all logs
berg ssh <inst_name> # ssh into an instance (then run `tt` to get tmux panes)
berg kill <inst_name> # kill an experiment
berg sync <inst_name> # sync local code to a running instance (for development)

You can also use the gce web interface which allows filtering and bulk management of all of your instances.

Running a large scale job with MPI

berg devbox -m <num_machines>

This will start num_machines devboxes with 4 GPUs each.

Then ssh into the first machine, and run mpi <your program> to run your program across all machines using MPI. Here mpi is an alias calling mpirun with all the necessary flags, and starting one worker process per GPU in your cluster. The easiest way of running multi machine tensorflow/pytorch experiments is to use Horovod, which now comes pre-installed on the golden image.

Creating a custom golden image

If you’d like to use a different image, you can edit the default_image value in ~/.berg/berg.json

We recommend starting with one of the official GCE deep learning images, and then modifying it slightly.

Berg will log into your image as the root user, and requires the following:

python installed with conda in /root/anaconda3/bin
A directory of code repositories in /root/code
Berg cloned to /root/code/berg and installed with pip install -e ~/code/berg

google/cloud-berg