Analyzing workflow cost (cloud only)

Note:

This tutorial assumes that you have access to Arvados command line tools, configured your API token, and confirmed a working environment.

Note:

This is only applicable when Arvados runs in a cloud environment and arvados-dispatch-cloud is used to dispatch crunch jobs. The per node-hour price for each defined InstanceType most be supplied in config.yml.

The arvados-client program can be used to analyze the cost of a workflow. It can be installed from packages (apt install arvados-client or yum install arvados-client). The arvados-client costanalyzer command analyzes the cost accounting information associated with Arvados container requests.

Syntax

The arvados-client costanalyzer tool has a number of command line arguments:

~$ arvados-client costanalyzer -h
Usage:
  ./arvados-client costanalyzer [options ...] [UUID ...]

  This program analyzes the cost of Arvados container requests and calculates
  the total cost across all requests. At least one UUID or a timestamp range
  must be specified.

  When the '-output' option is specified, a set of CSV files with cost details
  will be written to the provided directory. Each file is a CSV report that lists
  all the containers used to fulfill the container request, together with the
  machine type and cost of each container.

  When supplied with the UUID of a container request, it will calculate the
  cost of that container request and all its children.

  When supplied with the UUID of a collection, it will see if there is a
  container_request UUID in the properties of the collection, and if so, it
  will calculate the cost of that container request and all its children.

  When supplied with a project UUID or when supplied with multiple container
  request or collection UUIDs, it will calculate the total cost for all
  supplied UUIDs.

  When supplied with a 'begin' and 'end' timestamp (format:
  2006-01-02T15:04:05), it will calculate the cost for all top-level container
  requests whose containers finished during the specified interval.

  The total cost calculation takes container reuse into account: if a container
  was reused between several container requests, its cost will only be counted
  once.

  Caveats:

  - This program uses the cost data from config.yml at the time of the
  execution of the container, stored in the 'node.json' file in its log
  collection. If the cost data was not correctly configured at the time the
  container was executed, the output from this program will be incorrect.

  - If a container was run on a preemptible ("spot") instance, the cost data
  reported by this program may be wildly inaccurate, because it does not have
  access to the spot pricing in effect for the node then the container ran. The
  UUID report file that is generated when the '-output' option is specified has
  a column that indicates the preemptible state of the instance that ran the
  container.

  - This program does not take into account overhead costs like the time spent
  starting and stopping compute nodes that run containers, the cost of the
  permanent cloud nodes that provide the Arvados services, the cost of data
  stored in Arvados, etc.

  - When provided with a project UUID, subprojects will not be considered.

  In order to get the data for the UUIDs supplied, the ARVADOS_API_HOST and
  ARVADOS_API_TOKEN environment variables must be set.

  This program prints the total dollar amount from the aggregate cost
  accounting across all provided UUIDs on stdout.

Options:
  -begin begin
      timestamp begin for date range operation (format: 2006-01-02T15:04:05)
  -cache
      create and use a local disk cache of Arvados objects (default true)
  -end end
      timestamp end for date range operation (format: 2006-01-02T15:04:05)
  -log-level level
      logging level (debug, info, ...) (default "info")
  -output directory
      output directory for the CSV reports

Previous: Debugging workflows - shell access Next: Federated Multi-Cluster Workflows