Merge branch '23078-close-deleted-files'

[arvados.git] / doc / user / cwl / cwl-extensions.html.textile.liquid

---
layout: default
navsection: userguide
title: Arvados CWL Extensions
...
{% comment %}
Copyright (C) The Arvados Authors. All rights reserved.

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

Arvados provides several extensions to CWL for workflow optimization, site-specific configuration, and to enable access the Arvados API.

To use Arvados CWL extensions, add the following @$namespaces@ section at the top of your CWL file:

{% codeblock as yaml %}
$namespaces:
  arv: "http://arvados.org/cwl#"
  cwltool: "http://commonwl.org/cwltool#"
{% endcodeblock %}

For portability, most Arvados extensions should go into the @hints@ section of your CWL file.  This makes it possible for your workflows to run other CWL runners that do not recognize Arvados hints.  The difference between @hints@ and @requirements@ is that @hints@ are optional features that can be ignored by other runners and still produce the same output, whereas @requirements@ will fail the workflow if they cannot be fulfilled.  For example, @arv:IntermediateOutput@ should go in @hints@ as it will have no effect on non-Arvados platforms, however if your workflow explicitly accesses the Arvados API and will fail without it, you should put @arv:APIRequirement@ in @requirements@.

* "RunInSingleContainer":#RunInSingleContainer
* "SeparateRunner":#SeparateRunner
* "RuntimeConstraints":#RuntimeConstraints
* "PartitionRequirement":#PartitionRequirement
* "APIRequirement":#APIRequirement
* "IntermediateOutput":#IntermediateOutput
* "Secrets":#Secrets
* "WorkflowRunnerResources":#WorkflowRunnerResources
* "ClusterTarget":#ClusterTarget
* "OutputStorageClass":#OutputStorageClass
* "ProcessProperties":#ProcessProperties
* "OutputCollectionProperties":#OutputCollectionProperties
* "CUDARequirement":#CUDARequirement
* "ROCmRequirement":#ROCmRequirement
* "UsePreemptible":#UsePreemptible
* "PreemptionBehavior":#PreemptionBehavior
* "OutOfMemoryRetry":#OutOfMemoryRetry

{% codeblock as yaml %}
hints:
  arv:RunInSingleContainer: {}

  arv:SeparateRunner:
    runnerProcessName: $(inputs.sample_id)

  arv:RuntimeConstraints:
    keep_cache: 123456
    outputDirType: keep_output_dir

  arv:PartitionRequirement:
    partition: dev_partition

  arv:APIRequirement: {}

  arv:IntermediateOutput:
    outputTTL: 3600

  cwltool:Secrets:
    secrets: [input1, input2]

  arv:WorkflowRunnerResources:
    ramMin: 2048
    coresMin: 2
    keep_cache: 512

  arv:ClusterTarget:
    cluster_id: clsr1
    project_uuid: clsr1-j7d0g-qxc4jcji7n4lafx

  arv:OutputStorageClass:
    intermediateStorageClass: fast_storage
    finalStorageClass: robust_storage

  arv:ProcessProperties:
    processProperties:
      property1: value1
      property2: $(inputs.value2)

  arv:OutputCollectionProperties:
    outputProperties:
      property1: value1
      property2: $(inputs.value2)

  cwltool:CUDARequirement:
    cudaVersionMin: "11.0"
    cudaComputeCapability: "9.0"
    cudaDeviceCountMin: 1
    cudaDeviceCountMax: 1
    cudaVram: 8000

  arv:ROCmRequirement:
    rocmDriverVersion: "6.2"
    rocmTarget: ["gfx1100", "gfx1103"]
    rocmDeviceCountMin: 1
    rocmDeviceCountMax: 1
    rocmVram: 8000

  arv:UsePreemptible:
    usePreemptible: true

  arv:PreemptionBehavior:
    resubmitNonPreemptible: true

  arv:OutOfMemoryRetry:
    memoryRetryMultiplier: 2
    memoryErrorRegex: "custom memory error"
{% endcodeblock %}

h2(#RunInSingleContainer). arv:RunInSingleContainer

Apply this to a workflow step that runs a subworkflow.  Indicates that all the steps of the subworkflow should run together in a single container and not be scheduled separately.  If you have a sequence of short-running steps (less than 1-2 minutes each) this enables you to avoid scheduling and data transfer overhead by running all the steps together at once.  To use this feature, @cwltool@ must be installed in the container image.

h2(#SeparateRunner). arv:SeparateRunner

Apply this to a workflow step that runs a subworkflow.  Indicates that Arvados should launch a new workflow runner to manage that specific subworkflow instance.  If used on a scatter step, each scatter item is launched separately.  Using this option has three benefits:

* Better organization in the "Subprocesses" table of the main workflow, including the ability to provide a custom name for the step
* When re-running a batch that has run before, an entire subworkflow may be reused as a unit, which is faster than determining reuse for each step.
* Significantly faster submit rate compared to invoking @arvados-cwl-runner@ to launch individual workflow instances separately.

The disadvantage of this option is that because it does launch an additional workflow runner, that workflow runner consumes more compute resources compared to having all the steps managed by a single runner.

table(table table-bordered table-condensed).
|_. Field |_. Type |_. Description |
|runnerProcessName|optional string|Name to assign to the subworkflow process.  May be an expression with an input context of the post-scatter workflow step invocation.|

h2(#RuntimeConstraints). arv:RuntimeConstraints

Set Arvados-specific runtime hints.

table(table table-bordered table-condensed).
|_. Field |_. Type |_. Description |
|keep_cache|int|Size of file data buffer for Keep mount in MiB. Default is 256 MiB. Increase this to reduce cache thrashing in situations such as accessing multiple large (64+ MiB) files at the same time, or performing random access on a large file.|
|outputDirType|enum|Preferred backing store for output staging.  If not specified, the system may choose which one to use.  One of *local_output_dir* or *keep_output_dir*|

*local_output_dir*: Use regular file system local to the compute node. There must be sufficient local scratch space to store entire output; specify this with @outdirMin@ of @ResourceRequirement@.  Files are batch uploaded to Keep when the process completes.  Most compatible, but upload step can be time consuming for very large files.

*keep_output_dir*: Use writable Keep mount.  Files are streamed to Keep as they are written.  Does not consume local scratch space, but does consume RAM for output buffers (up to 192 MiB per file simultaneously open for writing.)  Best suited to processes which produce sequential output of large files (non-sequential writes may produced fragmented file manifests).  Supports regular files and directories, does not support special files such as symlinks, hard links, named pipes, named sockets, or device nodes.|

h2(#PartitionRequirement). arv:PartitionRequirement

Select preferred compute partitions on which to run jobs.

table(table table-bordered table-condensed).
|_. Field |_. Type |_. Description |
|partition|string or array of strings||

h2(#APIRequirement). arv:APIRequirement

For CWL v1.1 scripts, if a step requires network access but not specifically access to the Arvados API server, prefer the standard feature "NetworkAccess":https://www.commonwl.org/v1.1/CommandLineTool.html#NetworkAccess .  In the future, these may be differentiated by whether ARVADOS_API_HOST and ARVADOS_API_TOKEN is injected into the container or not.

Indicates that process wants to access to the Arvados API.  Will be granted network access and have @ARVADOS_API_HOST@ and @ARVADOS_API_TOKEN@ set in the environment.  Tools which rely on the Arvados API being present should put @arv:APIRequirement@ in the @requirements@ section of the tool (rather than @hints@) to indicate that that it is not portable to non-Arvados CWL runners.

Use @arv:APIRequirement@ in @hints@ to enable general (non-Arvados-specific) network access for a tool.

h2(#IntermediateOutput). arv:IntermediateOutput

Specify desired handling of intermediate output collections.

table(table table-bordered table-condensed).
|_. Field |_. Type |_. Description |
|outputTTL|int|If the value is greater than zero, consider intermediate output collections to be temporary and should be automatically trashed. Temporary collections will be trashed @outputTTL@ seconds after creation.  A value of zero means intermediate output should be retained indefinitely (this is the default behavior).
Note: arvados-cwl-runner currently does not take workflow dependencies into account when setting the TTL on an intermediate output collection. If the TTL is too short, it is possible for a collection to be trashed before downstream steps that consume it are started.  The recommended minimum value for TTL is the expected duration of the entire workflow.|

h2(#Secrets). cwltool:Secrets

Indicate that one or more input parameters are "secret".  Must be applied at the top level Workflow.  Secret parameters are not stored in keep, are hidden from logs and API responses, and are wiped from the database after the workflow completes.

*Note: currently, workflows with secrets must be submitted on the command line using @arvados-cwl-runner@.  Workflows with secrets submitted through Workbench will not properly obscure the secret inputs.*

table(table table-bordered table-condensed).
|_. Field |_. Type |_. Description |
|secrets|array<string>|Input parameters which are considered "secret".  Must be strings.|

h2(#WorkflowRunnerResources). arv:WorkflowRunnerResources

Specify resource requirements for the workflow runner process (arvados-cwl-runner) that manages a workflow run.  Must be applied to the top level workflow.  Will also be set implicitly when using @--submit-runner-ram@ on the command line along with @--create-workflow@ or @--update-workflow@.  Use this to adjust the runner's allocation if the workflow runner is getting "out of memory" exceptions or being killed by the out-of-memory (OOM) killer.

table(table table-bordered table-condensed).
|_. Field |_. Type |_. Description |
|ramMin|int|RAM, in mebibytes, to reserve for the arvados-cwl-runner process. Default 1 GiB|
|coresMin|int|Number of cores to reserve to the arvados-cwl-runner process. Default 1 core.|
|keep_cache|int|Size of collection metadata cache for the workflow runner, in MiB.  Default 256 MiB.  Will be added on to the RAM request when determining node size to request.|

h2(#ClusterTarget). arv:ClusterTarget

Specify which Arvados cluster should execute a container or subworkflow, and the parent project for the container request.

table(table table-bordered table-condensed).
|_. Field |_. Type |_. Description |
|cluster_id|string|The five-character alphanumeric cluster id (uuid prefix) where a container or subworkflow will execute.  May be an expression.|
|project_uuid|string|The uuid of the project which will own container request and output of the container.  May be an expression.|

h2(#OutputStorageClass). arv:OutputStorageClass

Specify the "storage class":{{site.baseurl}}/user/topics/storage-classes.html to use for intermediate and final outputs.

table(table table-bordered table-condensed).
|_. Field |_. Type |_. Description |
|intermediateStorageClass|string or array of strings|The storage class for output of intermediate steps.  For example, faster "hot" storage.|
|finalStorageClass_uuid|string or array of strings|The storage class for the final output.  |

h2(#ProcessProperties). arv:ProcessProperties

Specify extra "properties":{{site.baseurl}}/api/methods.html#subpropertyfilters that will be set on container requests created by the workflow.  May be set on a Workflow or a CommandLineTool.  Setting custom properties on a container request simplifies queries to find the workflow run later on.

table(table table-bordered table-condensed).
|_. Field |_. Type |_. Description |
|processProperties|key-value map, or list of objects with the fields {propertyName, propertyValue}|The properties that will be set on the container request.  May include expressions that reference @$(inputs)@ of the current workflow or tool.|

h2(#OutputCollectionProperties). arv:OutputCollectionProperties

Specify custom "properties":{{site.baseurl}}/api/methods.html#subpropertyfilters that will be set on the output collection of the workflow step.

table(table table-bordered table-condensed).
|_. Field |_. Type |_. Description |
|outputProperties|key-value map, or list of objects with the fields {propertyName, propertyValue}|The properties that will be set on the output collection.  May include expressions that reference @$(inputs)@ of the current workflow or tool.|

h2(#CUDARequirement). cwltool:CUDARequirement

Request support for Nvidia CUDA GPU acceleration in the container.  Assumes that the CUDA runtime (SDK) is installed in the container, and the host will inject the CUDA driver libraries into the container (equal or later to the version requested).

table(table table-bordered table-condensed).
|_. Field |_. Type |_. Description |
|cudaVersionMin|string|Required.  The CUDA SDK version corresponding to the minimum driver version supported by the container (generally, the SDK version 'X.Y' the application was compiled against).|
|cudaComputeCapability|string|Required.  The minimum CUDA hardware capability (in 'X.Y' format) required by the application's PTX or C++ GPU code (will be JIT compiled for the available hardware).|
|cudaDeviceCountMin|integer|Minimum number of GPU devices to allocate on a single node. Required.|
|cudaDeviceCountMax|integer|Maximum number of GPU devices to allocate on a single node. Optional.  If not specified, same as @cudaDeviceCountMin@.|
|cudaVram|integer|Requested amount of VRAM per device, in mebibytes (2**20)|

h2(#ROCmRequirement). cwltool:ROCmRequirement

Request support for AMD ROCm GPU acceleration in the container.  Assumes that the ROCm runtime (SDK) is installed in the container, and the host will inject the AMD devices (@/dev/kfd@ and @/dev/dri/renderD*@) container.

table(table table-bordered table-condensed).
|_. Field |_. Type |_. Description |
|rocmDriverVersion|string|Required.  The ROCm SDK version corresponding to the minimum driver version supported by the container (generally, the SDK version 'X.Y' the application was compiled against).|
|rocmTarget|array of string|Required.  A list of one or more hardware targets (e.g. gfx1100) corresponding to the GPU architectures supported by the container.  Use @rocminfo@ to determine what hardware targets you have.  See also "Accelerator and GPU hardware specifications":https://rocm.docs.amd.com/en/latest/reference/gpu-arch-specs.html (use the column "LLVM target name") and "LLVM AMDGPU backend documentation":https://llvm.org/docs/AMDGPUUsage.html .|
|rocmDeviceCountMin|integer|Minimum number of GPU devices to allocate on a single node. Required.|
|rocmDeviceCountMax|integer|Maximum number of GPU devices to allocate on a single node. Optional.  If not specified, same as @rocmDeviceCountMin@.|
|rocmVram|integer|Requested amount of VRAM per device, in mebibytes (2**20)|

h2(#UsePreemptible). arv:UsePreemptible

Specify whether a workflow step should request preemptible (e.g. AWS Spot market) instances.  Such instances are generally cheaper, but can be taken back by the cloud provider at any time (preempted) causing the step to fail.  When this happens, Arvados will automatically re-try the step, up to the configuration value of @Containers.MaxRetryAttempts@ (default 3) times.

table(table table-bordered table-condensed).
|_. Field |_. Type |_. Description |
|usePreemptible|boolean|Required, true to opt-in to using preemptible instances, false to opt-out.|

h2(#PreemptionBehavior). arv:PreemptionBehavior

This option determines the behavior when @arvados-cwl-runner@ detects that a workflow step was cancelled because the preemptible (spot market) instance it was running on was reclaimed by the cloud provider.  If 'true', instead of the retry behavior described above in 'UsePreemptible', on the first failure the workflow step will be re-submitted with preemption disabled, so it will be scheduled to run on non-preemptible (on-demand) instances.

When preemptible instances are reclaimed, this is a signal that the cloud provider has restricted capacity for low priority preemptible instance.  As a result, the default behavior of turning around and rescheduling or launching on another preemptible instance has higher risk of being preempted a second or third time, spending more time and money but making no progress.  This option provides an alternate fallback behavior, by attempting to run the step on a preemptible instance the first time (saving money), but re-running the step as non-preemptible if the first attempt was preempted (ensuring continued progress).

This behavior applied to each step individually.  If a step is preempted, then successfully re-run as non-preemptible, it does not affect the behavior of the next step, which will first be launched as preemptible, and so forth.

table(table table-bordered table-condensed).
|_. Field |_. Type |_. Description |
|resubmitNonPreemptible|boolean|Required.  If true, then when a workflow step is cancelled because the instance was preempted, re-submit the step with preemption disabled.|

h2(#OutOfMemoryRetry). arv:OutOfMemoryRetry

Specify that when a workflow step appears to have failed because it did not request enough RAM, it should be re-submitted with more RAM.  Out of memory conditions are detected either by the container being unexpectedly killed (exit code 137) or by matching a pattern in the container's output (see @memoryErrorRegex@).  Retrying will increase the base RAM request by the value of @memoryRetryMultiplier@.  For example, if the original RAM request was 10 GiB and the multiplier is 1.5, then it will re-submit with 15 GiB.

Containers are only re-submitted once.  If it fails a second time after increasing RAM, then the worklow step will still fail.

Also note that expressions that use @$(runtime.ram)@ (such as dynamic command line parameters) are not reevaluated when the container is resubmitted.

table(table table-bordered table-condensed).
|_. Field |_. Type |_. Description |
|memoryRetryMultiplier|float|Optional, default value is 2.  The retry will multiply the base memory request by this factor to get the retry memory request.|
|memoryErrorRegex|string|Optional, a custom regex that, if found in the stdout, stderr or crunch-run logging of a program, will trigger a retry with greater RAM.  If not provided, the default pattern matches "out of memory" (with or without spaces), "memory error" (with or without spaces), "bad_alloc" and "container using over 90% of memory".|

h2. arv:dockerCollectionPDH

This is an optional extension field appearing on the standard @DockerRequirement@.  It specifies the portable data hash of the Arvados collection containing the Docker image.  If present, it takes precedence over @dockerPull@ or @dockerImageId@.

<pre>
requirements:
  DockerRequirement:
    dockerPull: "debian:10"
    arv:dockerCollectionPDH: "feaf1fc916103d7cdab6489e1f8c3a2b+174"
</pre>

h1. Deprecated extensions

The following extensions are deprecated because equivalent features are part of the CWL v1.1 standard.

{% codeblock as yaml %}
hints:
  cwltool:LoadListingRequirement:
    loadListing: shallow_listing
  arv:ReuseRequirement:
    enableReuse: false
  cwltool:TimeLimit:
    timelimit: 14400
{% endcodeblock %}

h2. cwltool:LoadListingRequirement

For CWL v1.1 scripts, this is deprecated in favor of "loadListing":https://www.commonwl.org/v1.1/CommandLineTool.html#CommandInputParameter or "LoadListingRequirement":https://www.commonwl.org/v1.1/CommandLineTool.html#LoadListingRequirement

In CWL v1.0 documents, the default behavior for Directory objects is to recursively expand the @listing@ for access by parameter references an expressions.  For directory trees containing many files, this can be expensive in both time and memory usage.  Use @cwltool:LoadListingRequirement@ to change the behavior for expansion of directory listings in the workflow runner.

table(table table-bordered table-condensed).
|_. Field |_. Type |_. Description |
|loadListing|string|One of @no_listing@, @shallow_listing@, or @deep_listing@|

*no_listing*: Do not expand directory listing at all.  The @listing@ field on the Directory object will be undefined.

*shallow_listing*: Only expand the first level of directory listing.  The @listing@ field on the toplevel Directory object will contain the directory contents, however @listing@ will not be defined on subdirectories.

*deep_listing*: Recursively expand all levels of directory listing.  The @listing@ field will be provided on the toplevel object and all subdirectories.

h2. arv:ReuseRequirement

For CWL v1.1 scripts, this is deprecated in favor of "WorkReuse":https://www.commonwl.org/v1.1/CommandLineTool.html#WorkReuse .

Enable/disable work reuse for current process.  Default true (work reuse enabled).

table(table table-bordered table-condensed).
|_. Field |_. Type |_. Description |
|enableReuse|boolean|Enable/disable work reuse for current process.  Default true (work reuse enabled).|

h2. cwltool:TimeLimit

For CWL v1.1 scripts, this is deprecated in favor of "ToolTimeLimit":https://www.commonwl.org/v1.1/CommandLineTool.html#ToolTimeLimit

Set an upper limit on the execution time of a CommandLineTool or ExpressionTool.  A tool execution which exceeds the time limit may be preemptively terminated and considered failed.  May also be used by batch systems to make scheduling decisions.

table(table table-bordered table-condensed).
|_. Field |_. Type |_. Description |
|timelimit|int|Execution time limit in seconds. If set to zero, no limit is enforced.|