---
layout: default
navsection: installguide
title: Install the cloud dispatcher
...
{% comment %}
Copyright (C) The Arvados Authors. All rights reserved.

SPDX-License-Identifier: CC-BY-SA-3.0
{% endcomment %}

{% include 'notebox_begin_warning' %}
@arvados-dispatch-cloud@ is only relevant for cloud installations. Skip this section if you are installing an on premises cluster that will spool jobs to Slurm or LSF.
{% include 'notebox_end' %}

# "Introduction":#introduction
# "Create compute node VM image":#create-image
# "Update config.yml":#update-config
# "Install arvados-dispatch-cloud":#install-packages
# "Start the service":#start-service
# "Restart the API server and controller":#restart-api
# "Confirm working installation":#confirm-working

h2(#introduction). Introduction

The cloud dispatch service is for running containers on cloud VMs. It works with Microsoft Azure and Amazon EC2; future versions will also support Google Compute Engine.

The cloud dispatch service can run on any node that can connect to the Arvados API service, the cloud provider's API, and the SSH service on cloud VMs.  It is not resource-intensive, so you can run it on the API server node.

More detail about the internal operation of the dispatcher can be found in the "architecture section":{{site.baseurl}}/architecture/dispatchcloud.html.

h2(#update-config). Update config.yml

h3. Configure CloudVMs

Add or update the following portions of your cluster configuration file, @config.yml@. Refer to "config.defaults.yml":{{site.baseurl}}/admin/config.html for information about additional configuration options. The @DispatchPrivateKey@ should be the *private* key generated in "Create a SSH keypair":install-compute-node.html#sshkeypair .

<notextile>
<pre><code>    Services:
      DispatchCloud:
        InternalURLs:
          "http://localhost:9006": {}
    Containers:
      CloudVMs:
        # BootProbeCommand is a shell command that succeeds when an instance is ready for service
        BootProbeCommand: "sudo systemctl status docker"

        <b># --- driver-specific configuration goes here --- see Amazon and Azure examples below ---</b>

      DispatchPrivateKey: |
        -----BEGIN RSA PRIVATE KEY-----
        MIIEpQIBAAKCAQEAqXoCzcOBkFQ7w4dvXf9B++1ctgZRqEbgRYL3SstuMV4oawks
        ttUuxJycDdsPmeYcHsKo8vsEZpN6iYsX6ZZzhkO5nEayUTU8sBjmg1ZCTo4QqKXr
        FJ+amZ7oYMDof6QEdwl6KNDfIddL+NfBCLQTVInOAaNss7GRrxLTuTV7HcRaIUUI
        jYg0Ibg8ZZTzQxCvFXXnjseTgmOcTv7CuuGdt91OVdoq8czG/w8TwOhymEb7mQlt
        lXuucwQvYgfoUgcnTgpJr7j+hafp75g2wlPozp8gJ6WQ2yBWcfqL2aw7m7Ll88Nd
        [...]
        oFyAjVoexx0RBcH6BveTfQtJKbktP1qBO4mXo2dP0cacuZEtlAqW9Eb06Pvaw/D9
        foktmqOY8MyctzFgXBpGTxPliGjqo8OkrOyQP2g+FL7v+Km31Xs61P8=
        -----END RSA PRIVATE KEY-----
    InstanceTypes:
      x1md:
        ProviderType: x1.medium
        VCPUs: 8
        RAM: 64GiB
        IncludedScratch: 64GB
        Price: 0.62
      x1lg:
        ProviderType: x1.large
        VCPUs: 16
        RAM: 128GiB
        IncludedScratch: 128GB
        Price: 1.23
</code></pre>
</notextile>

h3(#GPUsupport). NVIDIA GPU support

To specify instance types with NVIDIA GPUs, "the compute image must be built with CUDA support":install-compute-node.html#nvidia , and you must include an additional @CUDA@ section:

<notextile>
<pre><code>    InstanceTypes:
      g4dn:
        ProviderType: g4dn.xlarge
        VCPUs: 4
        RAM: 16GiB
        IncludedScratch: 125GB
        Price: 0.56
        CUDA:
          DriverVersion: "11.4"
          HardwareCapability: "7.5"
          DeviceCount: 1
</code></pre>
</notextile>

The @DriverVersion@ is the version of the CUDA toolkit installed in your compute image (in X.Y format, do not include the patchlevel).  The @HardwareCapability@ is the "CUDA compute capability of the GPUs available for this instance type":https://developer.nvidia.com/cuda-gpus.  The @DeviceCount@ is the number of GPU cores available for this instance type.

h3(#aws-ebs-autoscaler). EBS Autoscale configuration

See "Autoscaling compute node scratch space":install-compute-node.html#aws-ebs-autoscaler for details about compute image configuration.

The @Containers.InstanceTypes@ list should be modified so that all @AddedScratch@ lines are removed, and the @IncludedScratch@ value should be set to 5 TB. This way, the scratch space requirements will be met by all the defined instance type. For example:

<notextile><pre><code>    InstanceTypes:
      c5large:
        ProviderType: c5.large
        VCPUs: 2
        RAM: 4GiB
        IncludedScratch: 5TB
        Price: 0.085
      m5large:
        ProviderType: m5.large
        VCPUs: 2
        RAM: 8GiB
        IncludedScratch: 5TB
        Price: 0.096
...
</code></pre></notextile>

You will also need to create an IAM role in AWS with these permissions:

<notextile><pre><code>{
    "Statement": [
        {
            "Effect": "Allow",
            "Action": [
                "ec2:AttachVolume",
                "ec2:DescribeVolumeStatus",
                "ec2:DescribeVolumes",
                "ec2:DescribeTags",
                "ec2:ModifyInstanceAttribute",
                "ec2:DescribeVolumeAttribute",
                "ec2:CreateVolume",
                "ec2:DeleteVolume",
                "ec2:CreateTags"
            ],
            "Resource": "*"
        }
    ]
}
</code></pre></notextile>

Then set @Containers.CloudVMs.DriverParameters.IAMInstanceProfile@ to the name of the IAM role. This will make @arvados-dispatch-cloud@ pass an IAM instance profile to the compute nodes when they start up, giving them sufficient permissions to attach and grow EBS volumes.

h3. AWS Credentials for Local Keepstore on Compute node

When @Containers.LocalKeepBlobBuffersPerVCPU@ is non-zero, the compute node will spin up a local Keepstore service for direct storage access. If Keep is backed by S3, the compute node will need to be able to access the S3 bucket.

If the AWS credentials for S3 access are configured in @config.yml@ (i.e. @Volumes.DriverParameters.AccessKeyID@ and @Volumes.DriverParameters.SecretAccessKey@), these credentials will be made available to the local Keepstore on the compute node to access S3 directly and no further configuration is necessary.

Alternatively, if an IAM role is configured in @config.yml@ (i.e. @Volumes.DriverParameters.IAMRole@), the name of an instance profile that corresponds to this role ("often identical to the name of the IAM role":https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/iam-roles-for-amazon-ec2.html#ec2-instance-profile) must be configured in the @CloudVMs.DriverParameters.IAMInstanceProfile@ parameter.

*If you are also using EBS Autoscale feature, the role in IAMInstanceProfile must have both ec2 and s3 permissions.*

Finally, if @config.yml@ does not have @Volumes.DriverParameters.AccessKeyID@, @Volumes.DriverParameters.SecretAccessKey@ or @Volumes.DriverParameters.IAMRole@ defined, Keepstore uses the IAM role attached to the node, whatever it may be called. The @CloudVMs.DriverParameters.IAMInstanceProfile@ parameter must then still be configured with the name of a profile whose IAM role has permission to access the S3 bucket(s). That way, @arvados-dispatch-cloud@ can attach the IAM role to the compute node as it is created.

h3. Minimal configuration example for Amazon EC2

The <span class="userinput">ImageID</span> value is the compute node image that was built in "the previous section":install-compute-node.html#aws.

<notextile>
<pre><code>    Containers:
      CloudVMs:
        ImageID: <span class="userinput">ami-01234567890abcdef</span>
        Driver: ec2
        DriverParameters:
          # If you are not using an IAM role for authentication, specify access
          # credentials here. Otherwise, omit or set AccessKeyID and
          # SecretAccessKey to an empty value.
          AccessKeyID: XXXXXXXXXXXXXXXXXXXX
          SecretAccessKey: YYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYYY

          SecurityGroupIDs:
          - sg-0123abcd
          SubnetID: subnet-0123abcd
          Region: us-east-1
          EBSVolumeType: gp2
          AdminUsername: arvados
</code></pre>
</notextile>

h3(#IAM). Example IAM policy for cloud dispatcher

Example policy for the IAM role used by the cloud dispatcher:

<notextile>
<pre>
{
    "Id": "arvados-dispatch-cloud policy",
    "Statement": [
        {
            "Effect": "Allow",
            "Action": [
                  "ec2:CreateTags",
                  "ec2:Describe*",
                  "ec2:CreateImage",
                  "ec2:CreateKeyPair",
                  "ec2:ImportKeyPair",
                  "ec2:DeleteKeyPair",
                  "ec2:RunInstances",
                  "ec2:StopInstances",
                  "ec2:TerminateInstances",
                  "ec2:ModifyInstanceAttribute",
                  "ec2:CreateSecurityGroup",
                  "ec2:DeleteSecurityGroup",
                  "iam:PassRole"
            ],
            "Resource": "*"
        }
    ]
}
</pre>
</notextile>

h3. Minimal configuration example for Azure

Using managed disks:

The <span class="userinput">ImageID</span> value is the compute node image that was built in "the previous section":install-compute-node.html#azure.

<notextile>
<pre><code>    Containers:
      CloudVMs:
        ImageID: <span class="userinput">"zzzzz-compute-v1597349873"</span>
        Driver: azure
        # (azure) managed disks: set MaxConcurrentInstanceCreateOps to 20 to avoid timeouts, cf
        # https://docs.microsoft.com/en-us/azure/virtual-machines/linux/capture-image
        MaxConcurrentInstanceCreateOps: 20
        DriverParameters:
          # Credentials.
          SubscriptionID: XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX
          ClientID: XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX
          ClientSecret: XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
          TenantID: XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX

          # Data center where VMs will be allocated
          Location: centralus

          # The resource group where the VM and virtual NIC will be
          # created.
          ResourceGroup: zzzzz
          NetworkResourceGroup: yyyyy   # only if different from ResourceGroup
          Network: xxxxx
          Subnet: xxxxx-subnet-private

          # The resource group where the disk image is stored, only needs to
          # be specified if it is different from ResourceGroup
          ImageResourceGroup: aaaaa

</code></pre>
</notextile>

Azure recommends using managed images. If you plan to start more than 20 VMs simultaneously, Azure recommends using a shared image gallery instead to avoid slowdowns and timeouts during the creation of the VMs.

Using an image from a shared image gallery:

<notextile>
<pre><code>    Containers:
      CloudVMs:
        ImageID: <span class="userinput">"shared_image_gallery_image_definition_name"</span>
        Driver: azure
        DriverParameters:
          # Credentials.
          SubscriptionID: XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX
          ClientID: XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX
          ClientSecret: XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
          TenantID: XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX

          # Data center where VMs will be allocated
          Location: centralus

          # The resource group where the VM and virtual NIC will be
          # created.
          ResourceGroup: zzzzz
          NetworkResourceGroup: yyyyy   # only if different from ResourceGroup
          Network: xxxxx
          Subnet: xxxxx-subnet-private

          # The resource group where the disk image is stored, only needs to
          # be specified if it is different from ResourceGroup
          ImageResourceGroup: aaaaa

          # (azure) shared image gallery: the name of the gallery
          SharedImageGalleryName: "shared_image_gallery_1"
          # (azure) shared image gallery: the version of the image definition
          SharedImageGalleryImageVersion: "0.0.1"

</code></pre>
</notextile>

Using unmanaged disks (deprecated):

The <span class="userinput">ImageID</span> value is the compute node image that was built in "the previous section":install-compute-node.html#azure.

<notextile>
<pre><code>    Containers:
      CloudVMs:
        ImageID: <span class="userinput">"https://zzzzzzzz.blob.core.windows.net/system/Microsoft.Compute/Images/images/zzzzz-compute-osDisk.55555555-5555-5555-5555-555555555555.vhd"</span>
        Driver: azure
        DriverParameters:
          # Credentials.
          SubscriptionID: XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX
          ClientID: XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX
          ClientSecret: XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX
          TenantID: XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXXXXXX

          # Data center where VMs will be allocated
          Location: centralus

          # The resource group where the VM and virtual NIC will be
          # created.
          ResourceGroup: zzzzz
          NetworkResourceGroup: yyyyy   # only if different from ResourceGroup
          Network: xxxxx
          Subnet: xxxxx-subnet-private

          # Where to store the VM VHD blobs
          StorageAccount: example
          BlobContainer: vhds

</code></pre>
</notextile>

Get the @SubscriptionID@ and @TenantID@:

<pre>
$ az account list
[
  {
    "cloudName": "AzureCloud",
    "id": "XXXXXXXX-XXXX-XXXX-XXXX-XXXXXXXX",
    "isDefault": true,
    "name": "Your Subscription",
    "state": "Enabled",
    "tenantId": "YYYYYYYY-YYYY-YYYY-YYYYYYYY",
    "user": {
      "name": "you@example.com",
      "type": "user"
    }
  }
]
</pre>

You will need to create a "service principal" to use as a delegated authority for API access.

<notextile><pre><code>$ az ad app create --display-name "Arvados Dispatch Cloud (<span class="userinput">ClusterID</span>)" --homepage "https://arvados.org" --identifier-uris "https://<span class="userinput">ClusterID.example.com</span>" --end-date 2299-12-31 --password <span class="userinput">Your_Password</span>
$ az ad sp create "<span class="userinput">appId</span>"
(appId is part of the response of the previous command)
$ az role assignment create --assignee "<span class="userinput">objectId</span>" --role Owner --scope /subscriptions/{subscriptionId}/
(objectId is part of the response of the previous command)
</code></pre></notextile>

Now update your @config.yml@ file:

@ClientID@ is the 'appId' value.

@ClientSecret@ is what was provided as <span class="userinput">Your_Password</span>.

h3. Test your configuration

Run the @cloudtest@ tool to verify that your configuration works. This creates a new cloud VM, confirms that it boots correctly and accepts your configured SSH private key, and shuts it down.

<notextile>
<pre><code>~$ <span class="userinput">arvados-server cloudtest && echo "OK!"</span>
</code></pre>
</notextile>

Refer to the "cloudtest tool documentation":../../admin/cloudtest.html for more information.

{% assign arvados_component = 'arvados-dispatch-cloud' %}

{% include 'install_packages' %}

{% include 'start_service' %}

{% include 'restart_api' %}

h2(#confirm-working). Confirm working installation

On the dispatch node, start monitoring the arvados-dispatch-cloud logs:

<notextile>
<pre><code># <span class="userinput">journalctl -o cat -fu arvados-dispatch-cloud.service</span>
</code></pre>
</notextile>

In another terminal window, use the diagnostics tool to run a simple container.

<notextile>
<pre><code># <span class="userinput">arvados-client sudo diagnostics</span>
INFO       5: running health check (same as `arvados-server check`)
INFO      10: getting discovery document from https://zzzzz.arvadosapi.com/discovery/v1/apis/arvados/v1/rest
...
INFO     160: running a container
INFO      ... container request submitted, waiting up to 10m for container to run
</code></pre>
</notextile>

After performing a number of other quick tests, this will submit a new container request and wait for it to finish.

While the diagnostics tool is waiting, the @arvados-dispatch-cloud@ logs will show details about creating a cloud instance, waiting for it to be ready, and scheduling the new container on it.

You can also use the "arvados-dispatch-cloud API":{{site.baseurl}}/api/dispatch.html to get a list of queued and running jobs and cloud instances. Use your @ManagementToken@ to test the dispatcher's endpoint. For example, when one container is running:

<notextile>
<pre><code>~$ <span class="userinput">curl -sH "Authorization: Bearer $token" http://localhost:9006/arvados/v1/dispatch/containers</span>
{
  "items": [
    {
      "container": {
        "uuid": "zzzzz-dz642-hdp2vpu9nq14tx0",
        ...
        "state": "Running",
        "scheduling_parameters": {
          "partitions": null,
          "preemptible": false,
          "max_run_time": 0
        },
        "exit_code": 0,
        "runtime_status": null,
        "started_at": null,
        "finished_at": null
      },
      "instance_type": {
        "Name": "Standard_D2s_v3",
        "ProviderType": "Standard_D2s_v3",
        "VCPUs": 2,
        "RAM": 8589934592,
        "Scratch": 16000000000,
        "IncludedScratch": 16000000000,
        "AddedScratch": 0,
        "Price": 0.11,
        "Preemptible": false
      }
    }
  ]
}
</code></pre>
</notextile>

A similar request can be made to the @http://localhost:9006/arvados/v1/dispatch/instances@ endpoint.

After the container finishes, you can get the container record by UUID *from a shell server* to see its results:

<notextile>
<pre><code>shell:~$ <span class="userinput">arv get <b>zzzzz-dz642-hdp2vpu9nq14tx0</b></span>
{
 ...
 "exit_code":0,
 "log":"a01df2f7e5bc1c2ad59c60a837e90dc6+166",
 "output":"d41d8cd98f00b204e9800998ecf8427e+0",
 "state":"Complete",
 ...
}
</code></pre>
</notextile>

You can use standard Keep tools to view the container's output and logs from their corresponding fields.  For example, to see the logs from the collection referenced in the @log@ field:

<notextile>
<pre><code>~$ <span class="userinput">arv keep ls <b>a01df2f7e5bc1c2ad59c60a837e90dc6+166</b></span>
./crunch-run.txt
./stderr.txt
./stdout.txt
~$ <span class="userinput">arv-get <b>a01df2f7e5bc1c2ad59c60a837e90dc6+166</b>/stdout.txt</span>
2016-08-05T13:53:06.201011Z Hello, Crunch!
</code></pre>
</notextile>

If the container does not dispatch successfully, refer to the @arvados-dispatch-cloud@ logs for information about why it failed.