Deploy Determined with Dynamic Agents#

This document describes how to install, configure, and upgrade a deployment of Determined with dynamic agents on GCP. For an overview of the elastic infrastructure in Determined, visit Elastic Infrastructure.

System Requirements#

Compute Engine Project#

The Determined master and the Determined agents are intended to run in the same project.

Instance Labels#

When using dynamic agents on GCP, Determined identifies the Compute Engine instances that it is managing using a configurable instance label (see Cluster Configuration for details). Administrators should be careful to ensure that this label is not used by other Compute Engine instances that are launched outside of Determined; if that assumption is violated, unexpected behavior may occur.

Compute Engine Images#

  • The Determined master node will run on a custom image that will be shared with you by Determined AI.

  • Determined agent nodes will run on a custom image that will be shared with you by Determined AI.

Compute Engine Machine Types#

  • The Determined master node should be deployed on a Compute Engine instance with >= 2 CPUs (Intel Broadwell or later), 4GB of RAM, and 100GB of disk storage. This would be a Compute Engine n1-standard-2 or more powerful.

GCP API Access#

  • The Determined master needs to run as a service account that has the permissions to manage Compute Engine instances. There are two options:

    1. Create a particular service account with the Compute Admin role. Then set the Determined master to use this account. See Compute Engine IAM roles for more details on how to configure the service account.

      • In order for the Determined agent to be associated with a service account, the Determined master needs to have access to service accounts. Please ensure the service account of the Determined master has the Service Account User role.

      • In order for the Determined agent to use a shared VPC, the service account that the master runs with needs to have the Compute Network User role.

    2. Use the default service account and add the Compute Engine: Read Write scope.

  • Optionally, the Determined agent may be associated with a service account.


Access scopes are the legacy method of specifying permissions for your instance. A best practice is to set the full cloud-platform access scope on the instance, then securely limit the service account’s API access with Cloud IAM roles. See Access Scopes for details.

Set up Internet Access#

  • The Determined Docker images are hosted on Docker Hub. Determined agents need access to Docker Hub for such tasks as building new images for user workloads.

  • If packages, data, or other resources needed by user workloads are hosted on the public Internet, Determined agents need to be able to access them. Note that agents can be configured to use proxies when accessing network resources.

  • For best performance, it is recommended that the Determined master and agents use the same physical network or VPC. When using VPCs on a public cloud provider, additional steps might need to be taken to ensure that instances in the VPC can access the Internet:

    • On GCP, the instances need to have an external IP address, or a GCP Cloud NAT should be configured for the VPC.

    • On AWS, the instances need to have a public IP address, and a VPC Internet Gateway should be configured for the VPC.

Set up Firewall Rules#

The firewall rules must satisfy the following network access requirements for the master and agents.


  • Inbound TCP to the master’s network port from the Determined agent instances, as well as all machines where developers want to use the Determined CLI or WebUI. The default port is 8443 if TLS is enabled and 8080 if not.

  • Outbound TCP to all ports on the Determined agents.


  • Inbound TCP from all ports on the master to all ports on the agent.

  • Outbound TCP from all ports on the agent to the master’s network port.

  • Outbound TCP to the services that host the Docker images, packages, data, and other resources that need to be accessed by user workloads.

    • For example, if your data is stored on Amazon S3, ensure the firewall rules allow access to this data.

  • Inbound and outbound TCP on all ports to and from each Determined agent. The details are as follows:

    • Inbound and outbound TCP ports 1734 and 1750 are used for synchronization between trial containers.

    • Inbound and outbound TCP port 12350 is used for internal SSH-based communication between trial containers.

    • Inbound and outbound TCP port 12355 is used for GLOO rendezvous between trial containers.

    • Inbound and outbound ephemeral TCP ports in the range 1024-65536 are used for communication between trials via GLOO. This range is configured by the configuration parameter task_container_defaults.gloo_port_range inside master.yaml as described in the Configuring the Cluster guide.

    • For every GPU on each agent machine, an inbound and outbound ephemeral TCP port in the range 1024-65536 is used for communication between trials via NCCL. This range is configured by the configuration parameter task_container_defaults.nccl_port_range inside master.yaml as described in the Configuring the Cluster guide.

    • Two additional ephemeral TCP ports in the range 1024-65536 are used for additional intra-trial communication between trial containers.

    • For TensorBoards, an inbound and outbound TCP port between 2600-2900 is used to connect the master and the tensorboard container.

The following GPU types are supported by Determined:

  • nvidia-tesla-t4

  • nvidia-tesla-p100

  • nvidia-tesla-p4

  • nvidia-tesla-v100

  • nvidia-tesla-a100

Cluster Configuration#

The Determined Cluster is configured with master.yaml file located at /usr/local/determined/etc on the Determined master instance. You need to configure GPU dynamic agents in each resource pool. See Configuring the Cluster for details.

Attach a Disk To Each Dynamic Agent#

If your input data set is on a persistent disk, you can attach that disk to each dynamic agent by using the base instance configuration and preparing commands. The following is an example configuration. See REST Resource: instances for the full list of configuration options supported by GCP. See Formatting and mounting a zonal persistent disk for more examples of formatting or mounting disks in GCP.

Here is an example master configuration to attach and mount a second disk to each dynamic agent.

  startup_script: |
                  mkdir -p /mnt/disks/second
                  mount -o discard,defaults /dev/sdb1 /mnt/disks/second
  type: gcp
      - mode: READ_ONLY
        boot: false
        source: zones/<zone>/disks/<the name of the existing disk>
        autoDelete: false
  boot_disk_size: 200
  boot_disk_source_image: projects/<project>/global/images/<image name>


If a specific non-root user needs to access the disk, please run the tasks linked with the POSIX UID/GID of the user (See Run Tasks as Specific Agent Users for details.) and grant access to the corresponding UID/GID.

You can use the following command to validate if Determined tasks can read from the attached disk.

cat > command.yaml << EOF
  - host_path: /mnt/disks/second
    container_path: /second
# Test attached read-only disk.
det command run --config-file command.yaml ls -l /second

Securely Pull Images from GCR#

If you have expensive operations to perform at startup, it can be useful to add custom layers to the task images Determined provides. If you have store these images in a secure registry, such as GCR, you can pull these images securely by using existing tooling like docker-credential-gcr.

Here is an example master configuration of how to allow the agent to inherit the permissions of the service account associated with a GCE instance, for accessing GCR.

  container_startup_script: |
      export HOME=/root
      apt-get update && apt-get install -y curl
      curl -fsSL "" \
            | tar xz --to-stdout > /usr/bin/docker-credential-gcr && chmod +x /usr/bin/docker-credential-gcr
      docker-credential-gcr configure-docker


This is an example of an operation that requires use of container_startup_script. Because Docker credential helpers alter the Docker client configuration to depend on the helper binary by name, it must be installed and configured in the container.


These instructions describe how to install Determined for the first time; for directions on how to upgrade an existing Determined installation, see the Upgrades section below.

Ensure that you are using the most up-to-date Determined images. Keep the image IDs handy as we will need them later.


To install the master, we will launch an instance from the Determined master image.

Let’s start by navigating to the Compute Engine Dashboard of the GCP Console. Click “Create Instance” and follow the instructions below:

  1. Choose Machine Type: we recommend a n1-standard-2 or more powerful.

  2. Configure Boot Disk:

    1. Choose Boot Disk Image: find the Determined master image in “Images” and click “Select”.

    2. Set Boot Disk Size: set Size to be at least 100GB. If you have a previous Determined installation that you are upgrading, you want to use the snapshot or existing disk. This disk will be used to store all your experiment metadata and checkpoints.

  3. Configure Identity and API access: choose the service account according to GCP API Access.

  4. Configure Firewalls: choose or create a security group according to these Set up Internet Access. Check off Allow HTTP traffic.

  5. Review and launch the instance.

  6. SSH into the Determined master and edit the config at /usr/local/determined/etc/master.yaml according to the guide on Configuring the Cluster.

  7. Start the Determined master by entering make -C /usr/local/determined enable-master into the terminal.


There is no installation needed for the agent. The Determined master will dynamically launch Determined agent instances based on the Configuring the Cluster.


Upgrading an existing Determined installation with dynamic agents on GCP requires the same steps as an installation without dynamic agents. See Upgrade.