cloudman 0.0.8

Command line tool for managing single-purpose cloud VMs

cloudman

Command line tool for managing single-purpose cloud VMs. Currently works with GCP.

Features:

1. Single command to provision cloud-based disk images with Ubuntu, NVIDIA drivers, Anaconda, PyTorch, FastAI and more..

2. Single command to attach a disk image, launch a cloud VM instance with any configuration of CPUs, RAM and GPU, and start Jupyter notebook.

Installation

Install the cloudman package using pip.

pip install cloudman

cloudman requires the gcloud command line tool to be installed. Make sure to complete these steps before moving forward.

1. Install gcloud CLI using this link: https://cloud.google.com/sdk/docs/downloads-interactive

2. Make sure the gcloud command is added to your PATH.

3. Connect the gcloud tool your GCP account by running:

   gcloud init

4. (Optional) Create a GCP project by running:

   gcloud projects create PROJECT_ID

Replace PROJECT_ID with a unique name e.g. kitten-puppies-999 You can also use an existing project.

5. Activate a project by running:

   gcloud config set project PROJECT_ID

6. Open the cloud console by running:

   cloudman console

7. Enable billing to start creating cloud VMs.

Usage

cloudman allows you to crate standalone boot disks which can be flexibly attached to any type of machine & GPU configuration.

Create a boot disk

Running the cloudman create will create a boot disk of the given size, then install Ubuntu, NVIDIA CUDA libraries, Anaconda, Pytorch and FastAI deep learning libraries. By default, a 50 GB disk is created.

Here are some examples:

# Using default disk size of 50GB
cloudman create freesound-2019

# Custom disk size
cloudman create freesound-2019 --disk=100GB

Note that cloudman creates a temporary VM instance to install all the required packages, and then deletes the VM, but retains the disk. It uses this shell script for setup.

Start an instance

You can launch an instance with the boot disk attached using the cloudman start command. You can choose any combination of GPUs and machine types. You can also launch a CPU-only instance, which is the default. Also, you can choose whether you want a preemptible instance (enabled by default, to save cost), or a dedicated instance.

Here are some examples:

# CPU-only, preemptible
cloudman start jigsaw-ulmfit

# Same as the above command, but explicit
cloudman start jigsaw-ulmfit --gpu=nogpu --machine=auto --spot

# K80 GPU, dedicated
cloudman start jigsaw-ulmfit --gpu=k80 --nospot

Valid GPU types are: 'nogpu', 't4', 'v100', 'p100', 'p4', 'k80' Valid machine types can be found here: https://cloud.google.com/compute/docs/machine-types If machine type is set to 'auto', the tool automatically picks a reasonable machine type to match the GPU's RAM & cores. You can see the mapping here.

Once started, you can use the link to access the Jupyter server. Sometimes it may take a minute or two for the Jupyter server to start.

SSH into running instance

To SSH into a running instance use:

cloudman ssh jigsaw-ulmfit

Stop an instance

Use the cloudman stop command to stop a running instance. Example:

cloudman stop jigsaw-ulmfit

List disks and running instances

To see the available boot disks and running instances, run

cloudman list

Delete a boot disk

To delete a boot disk, run

cloudman delete jigsaw-ulmfit

This will stop any running instances and delte the boot disk complete. All your data will be lost, so be careful while using this.