Microsoft Azure

AzureVMCluster([location, resource_group, ...])

Cluster running on Azure Virtual machines.

Overview

Authentication

In order to create clusters on Azure you need to set your authentication credentials. You can do this via the az command line tool.

$ az login

Note

Setting the default output to table with az configure will make the az tool much easier to use.

Resource Groups

To create resources on Azure they must be placed in a resource group. Dask Cloudprovider will need a group to create Dask components in.

You can list existing groups via the cli.

$ az group list

You can also create a new resource group if you do not have an existing one.

$ az group create --location <location> --name <resource group name> --subscription <subscription>

You can get a full list of locations with az account list-locations and subscriptions with az account list.

Take note of your resource group name for later.

Virtual Networks

Compute resources on Azure must be placed in virtual networks (vnet). Dask Cloudprovider will require an existing vnet to connect compute resources to.

You can list existing vnets via the cli.

$ az network vnet list

You can also create a new vnet via the cli.

$ az network vnet create -g <resource group name> -n <vnet name> --address-prefix 10.0.0.0/16 \
      --subnet-name <subnet name> --subnet-prefix 10.0.0.0/24

This command will create a new vnet in your resource group with one subnet with the 10.0.0.0/24 prefix. For more than 255 compute resources you will need additional subnets.

Take note of your vnet name for later.

Security Groups

To allow network traffic to reach your Dask cluster you will need to create a security group which allows traffic on ports 8786-8787 from wherever you are.

You can list existing security groups via the cli.

$ az network nsg list

Or you can create a new security group.

$ az network nsg create -g <resource group name> --name <security group name>
$ az network nsg rule create -g <resource group name> --nsg-name <security group name> -n MyNsgRuleWithAsg \
      --priority 500 --source-address-prefixes Internet --destination-port-ranges 8786 8787 \
      --destination-address-prefixes '*' --access Allow --protocol Tcp --description "Allow Internet to Dask on ports 8786,8787."

This example allows all traffic to 8786-8787 from the internet. It is recommended you make your rules more restrictive than this by limiting it to your corporate network or specific IP.

Again take note of this security group name for later.

Dask Configuration

You’ll provide the names or IDs of the Azure resources when you create a AzureVMCluster. You can specify these values manually, or use Dask’s configuration system system. For example, the resource_group value can be specified using an environment variable:

$ export DASK_CLOUDPROVIDER__AZURE__RESOURCE_GROUP="<resource group name>"
$ python

Or you can set it in a YAML configuration file.

cloudprovider:
  azure:
    resource_group: "<resource group name>"
    azurevm:
     vnet: "<vnet name>"

Note that the options controlling the VMs are under the cloudprovider.azure.azurevm key.

See Configuration for more.

AzureVM

class dask_cloudprovider.azure.AzureVMCluster(location: Optional[str] = None, resource_group: Optional[str] = None, vnet: Optional[str] = None, security_group: Optional[str] = None, public_ingress: Optional[bool] = None, vm_size: Optional[str] = None, scheduler_vm_size: Optional[str] = None, vm_image: dict = {}, disk_size: Optional[int] = None, bootstrap: Optional[bool] = None, auto_shutdown: Optional[bool] = None, docker_image=None, debug: bool = False, marketplace_plan: dict = {}, subscription_id: Optional[str] = None, **kwargs)[source]

Cluster running on Azure Virtual machines.

This cluster manager constructs a Dask cluster running on Azure Virtual Machines.

When configuring your cluster you may find it useful to install the az tool for querying the Azure API for available options.

https://docs.microsoft.com/en-us/cli/azure/install-azure-cli

Parameters
location: str

The Azure location to launch you cluster in. List available locations with az account list-locations.

resource_group: str

The resource group to create components in. List your resource groups with az group list.

vnet: str

The vnet to attach VM network interfaces to. List your vnets with az network vnet list.

security_group: str

The security group to apply to your VMs. This must allow ports 8786-8787 from wherever you are running this from. List your security groups with az network nsg list.

public_ingress: bool

Assign a public IP address to the scheduler. Default True.

vm_size: str

Azure VM size to use for scheduler and workers. Default Standard_DS1_v2. List available VM sizes with az vm list-sizes --location <location>.

disk_size: int

Specifies the size of the VM host OS disk in gigabytes. Default is 50. This value cannot be larger than 1023.

scheduler_vm_size: str

Azure VM size to use for scheduler. If not set will use the vm_size.

vm_image: dict

By default all VMs will use the latest Ubuntu LTS release with the following configuration

{"publisher": "Canonical", "offer": "UbuntuServer","sku": "18.04-LTS", "version": "latest"}

You can override any of these options by passing a dict with matching keys here. For example if you wish to try Ubuntu 19.04 you can pass {"sku": "19.04"} and the publisher, offer and version will be used from the default.

bootstrap: bool (optional)

It is assumed that the VHD will not have Docker installed (or the NVIDIA drivers for GPU instances). If bootstrap is True these dependencies will be installed on instance start. If you are using a custom VHD which already has these dependencies set this to False.

auto_shutdown: bool (optional)

Shutdown the VM if the Dask process exits. Default True.

worker_module: str

The Dask worker module to start on worker VMs.

n_workers: int

Number of workers to initialise the cluster with. Defaults to 0.

worker_module: str

The Python module to run for the worker. Defaults to distributed.cli.dask_worker

worker_options: dict

Params to be passed to the worker class. See distributed.worker.Worker for default worker class. If you set worker_module then refer to the docstring for the custom worker class.

scheduler_options: dict

Params to be passed to the scheduler class. See distributed.scheduler.Scheduler.

docker_image: string (optional)

The Docker image to run on all instances.

This image must have a valid Python environment and have dask installed in order for the dask-scheduler and dask-worker commands to be available. It is recommended the Python environment matches your local environment where AzureVMCluster is being created from.

For GPU instance types the Docker image much have NVIDIA drivers and dask-cuda installed.

By default the daskdev/dask:latest image will be used.

docker_args: string (optional)

Extra command line arguments to pass to Docker.

extra_bootstrap: list[str] (optional)

Extra commands to be run during the bootstrap phase.

silence_logs: bool

Whether or not we should silence logging when setting up the cluster.

asynchronous: bool

If this is intended to be used directly within an event loop with async/await

securitySecurity or bool, optional

Configures communication security in this cluster. Can be a security object, or True. If True, temporary self-signed credentials will be created automatically. Default is True.

debug: bool, optional

More information will be printed when constructing clusters to enable debugging.

marketplace_plan: dict (optional)

Plan information dict necessary for creating a virtual machine from Azure Marketplace image or a custom image sourced from a Marketplace image with a plan. Default is {}.

All three fields “name”, “publisher”, “product” must be passed in the dictionary if set. For e.g.

{"name": "ngc-base-version-21-02-2", "publisher": "nvidia","product": "ngc_azure_17_11"}

subscription_id: str (optional)

The ID of the Azure Subscription to create the virtual machines in. If not specified, then dask-cloudprovider will attempt to use the configured default for the Azure CLI. List your subscriptions with az account list.

Examples

Minimal example

Create the cluster

>>> from dask_cloudprovider.azure import AzureVMCluster
>>> cluster = AzureVMCluster(resource_group="<resource group>",
...                          vnet="<vnet>",
...                          security_group="<security group>",
...                          n_workers=1)
Creating scheduler instance
Assigned public IP
Network interface ready
Creating VM
Created VM dask-5648cc8b-scheduler
Waiting for scheduler to run
Scheduler is running
Creating worker instance
Network interface ready
Creating VM
Created VM dask-5648cc8b-worker-e1ebfc0e

Connect a client.

>>> from dask.distributed import Client
>>> client = Client(cluster)

Do some work.

>>> import dask.array as da
>>> arr = da.random.random((1000, 1000), chunks=(100, 100))
>>> arr.mean().compute()
0.5004117488368686

Close the cluster.

>>> client.close()
>>> cluster.close()
Terminated VM dask-5648cc8b-worker-e1ebfc0e
Removed disks for VM dask-5648cc8b-worker-e1ebfc0e
Deleted network interface
Terminated VM dask-5648cc8b-scheduler
Removed disks for VM dask-5648cc8b-scheduler
Deleted network interface
Unassigned public IP

You can also do this all in one go with context managers to ensure the cluster is created and cleaned up.

>>> with AzureVMCluster(resource_group="<resource group>",
...                     vnet="<vnet>",
...                     security_group="<security group>",
...                     n_workers=1) as cluster:
...     with Client(cluster) as client:
...             print(da.random.random((1000, 1000), chunks=(100, 100)).mean().compute())
Creating scheduler instance
Assigned public IP
Network interface ready
Creating VM
Created VM dask-1e6dac4e-scheduler
Waiting for scheduler to run
Scheduler is running
Creating worker instance
Network interface ready
Creating VM
Created VM dask-1e6dac4e-worker-c7c4ca23
0.4996427609642539
Terminated VM dask-1e6dac4e-worker-c7c4ca23
Removed disks for VM dask-1e6dac4e-worker-c7c4ca23
Deleted network interface
Terminated VM dask-1e6dac4e-scheduler
Removed disks for VM dask-1e6dac4e-scheduler
Deleted network interface
Unassigned public IP

RAPIDS example

You can also use AzureVMCluster to run a GPU enabled cluster and leverage the RAPIDS accelerated libraries.

>>> cluster = AzureVMCluster(resource_group="<resource group>",
...                          vnet="<vnet>",
...                          security_group="<security group>",
...                          n_workers=1,
...                          vm_size="Standard_NC12s_v3",  # Or any NVIDIA GPU enabled size
...                          docker_image="rapidsai/rapidsai:cuda11.0-runtime-ubuntu18.04-py3.8",
...                          worker_class="dask_cuda.CUDAWorker")
>>> from dask.distributed import Client
>>> client = Client(cluster)

Run some GPU code.

>>> def get_gpu_model():
...     import pynvml
...     pynvml.nvmlInit()
...     return pynvml.nvmlDeviceGetName(pynvml.nvmlDeviceGetHandleByIndex(0))
>>> client.submit(get_gpu_model).result()
b'Tesla V100-PCIE-16GB'

Close the cluster.

>>> client.close()
>>> cluster.close()
Attributes
asynchronous

Are we running in the event loop?

auto_shutdown
bootstrap
command
dashboard_link
docker_image
gpu_instance
loop
name
observed
plan
requested
scheduler_address
scheduler_class
worker_class

Methods

adapt([Adaptive, minimum, maximum, ...])

Turn on adaptivity

call_async(f, *args, **kwargs)

Run a blocking function in a thread as a coroutine.

from_name(name)

Create an instance of this class to represent an existing cluster by name.

get_client()

Return client for the cluster

get_logs([cluster, scheduler, workers])

Return logs for the cluster, scheduler and workers

get_tags()

Generate tags to be applied to all resources.

new_worker_spec()

Return name and spec for the next worker

scale([n, memory, cores])

Scale cluster to n workers

scale_up([n, memory, cores])

Scale cluster to n workers

sync(func, *args[, asynchronous, ...])

Call func with args synchronously or asynchronously depending on the calling context

close

get_cloud_init

logs

render_cloud_init

render_process_cloud_init

scale_down

Azure Spot Instance Plugin

class dask_cloudprovider.azure.AzurePreemptibleWorkerPlugin(poll_interval_s=1, metadata_url=None, termination_events=None, termination_offset_minutes=0)[source]

A worker plugin for azure spot instances

This worker plugin will poll azure’s metadata service for preemption notifications. When a node is preempted, the plugin will attempt to shutdown gracefully all workers on the node.

This plugin can be used on any worker running on azure spot instances, not just the ones created by dask-cloudprovider.

For more details on azure spot instances see: https://docs.microsoft.com/en-us/azure/virtual-machines/linux/scheduled-events

Parameters
poll_interval_s: int (optional)

The rate at which the plugin will poll the metadata service in seconds.

Defaults to 1

metadata_url: str (optional)

The url of the metadata service to poll.

Defaults to “http://169.254.169.254/metadata/scheduledevents?api-version=2019-08-01

termination_events: List[str] (optional)

The type of events that will trigger the gracefull shutdown

Defaults to ['Preempt', 'Terminate']

termination_offset_minutes: int (optional)

Extra offset to apply to the premption date. This may be negative, to start the gracefull shutdown before the NotBefore date. It can also be positive, to start the shutdown after the NotBefore date, but this is at your own risk.

Defaults to 0

Examples

Let’s say you have cluster and a client instance. For example using dask_kubernetes.KubeCluster

>>> from dask_kubernetes import KubeCluster
>>> from distributed import Client
>>> cluster = KubeCluster()
>>> client = Client(cluster)

You can add the worker plugin using the following:

>>> from dask_cloudprovider.azure import AzurePreemptibleWorkerPlugin
>>> client.register_worker_plugin(AzurePreemptibleWorkerPlugin())

Methods

setup(worker)

Run when the plugin is attached to a worker.

teardown(worker)

Run when the worker to which the plugin is attached to is closed

transition(key, start, finish, **kwargs)

Throughout the lifecycle of a task (see Worker), Workers are instructed by the scheduler to compute certain tasks, resulting in transitions in the state of each task.

poll_status

setup(worker)[source]

Run when the plugin is attached to a worker. This happens when the plugin is registered and attached to existing workers, or when a worker is created after the plugin has been registered.

teardown(worker)[source]

Run when the worker to which the plugin is attached to is closed