Skip to content

Latest commit

 

History

History
141 lines (105 loc) · 4.9 KB

job_control.md

File metadata and controls

141 lines (105 loc) · 4.9 KB
Raw

Job Control

dsub provides two simple mechanisms for specifying job dependencies. This enables you to sequence or run in parallel multiple jobs, each of which can have its own resource requirements, such as CPU and memory.

Example with --wait

If your job dependencies are simple - one job needs to run after another - you can simply use the "--wait" flag for each job that you launch.

The "--wait" flag will cause dsub to block after launching your job. It will continue to poll until your job completes. If the job completes successfully, then dsub will exit with status code 0, otherwise dsub exits with status code 1.

You can take advantage of the "errexit" option in bash:

# Enable exit on error
set -o errexit

# Launch step 1 and block until completion
dsub ... --wait

# Launch step 2 and block until completion
dsub ... --wait

# Launch step 3
dsub ...

If you want to handle the error explicitly, the dsub exit status can be handled as:

# Launch step 1 and block until completion
if ! dsub ... --wait; then
  echo "Step 1 failed!"
  exit 1
fi

Example with --after

Suppose you want to run jobs A and B in parallel, and then job C.

This can be done with --after, passing it the job-id from A and B.

The first step is to capture the job-id for the job you want to wait for.

JOBID_A=$(dsub ...)
JOBID_B=$(dsub ...)

Then pass these values to the --after flag for launching Job C

dsub ... --after "${JOB_A}" "${JOB_B}"

Here is the output of a sample run:

$ JOBID_A=$(dsub --provider google-cls-v2 --project "${MYPROJECT}" --regions us-central1 \
--logging "gs://${MYBUCKET}/logging/"   \
--command 'echo "hello from job A"')
Job: echo--<user>--180924-112256-64
Launched job-id: echo--<user>--180924-112256-64
To check the status, run:
  dstat --provider google-cls-v2 --project ${MYPROJECT} --jobs 'echo--<user>--180924-112256-64' --status '*'
To cancel the job, run:
  ddel --provider google-cls-v2 --project ${MYPROJECT} --jobs 'echo--<user>--180924-112256-64'

$ echo "${JOBID_A}"
echo--<user>--180924-112256-64

$ JOBID_B=... (similar)

$ JOBID_C=$(dsub --provider google-cls-v2 --project "${MYPROJECT}" --regions us-central1 \
--logging "gs://${MYBUCKET}/logging/"   \
--command 'echo "job C"' --after "${JOBID_A}" "${JOBID_B}")
Waiting for predecessor jobs to complete...
Waiting for: echo--<user>--180924-112256-64, echo--<user>--180924-112259-48.
  echo--<user>--180924-112256-64: SUCCESS
Waiting for: echo--<user>--180924-112259-48.
  echo--<user>--180924-112259-48: SUCCESS
Launched job-id: echo--<user>--180924-112302-87
To check the status, run:
  dstat --provider google-cls-v2 --project ${MYPROJECT} --jobs 'echo--<user>--180924-112302-87' --status '*'
To cancel the job, run:
  ddel --provider google-cls-v2 --project ${MYPROJECT} --jobs 'echo--<user>--180924-112302-87'
echo--<user>--180924-112302-87

--after is blocking

The --after command works by blocking dsub until the previous job(s) complete. This means that the submission will fail if you turn off your computer before that happens.

If you'd like to be able to turn off your machine (perhaps it's a laptop and you're running a big job) then what you can do is put your dsub command itself into a script, and run that script using dsub itself.

Using --skip to bypass jobs that have already run

With the --skip parameter, dsub will skip running a job if all the outputs for the job already exist. This useful when you are building and debugging a sequence of jobs, such as:

JOBID_A=$(dsub ... --skip --output gs://${MYBUCKET}/a_file)
JOBID_B=$(dsub ... --skip --output gs://${MYBUCKET}/b_file)

dsub ... --after "${JOB_A}" "${JOB_B}"

If on your first run of this script, the first job, "job A" fails and the second job, "job B" succeeds, then when you fix "job A" and re-run the script, "job B" will be skipped. Only "job A" will be re-run. If it succeeds, then the third job will run.

The special NO_JOB return value

When a job is skipped because the output already exists, dsub will output a special job-id value, NO_JOB. When NO_JOB is passed to --after, dsub treats that job as completed successfully.

--skip caveats: wildcards and recursive output

When wildcards are used for --output parameters or --output-recursive parameters are used, there is no way for dsub to verify that all output is present. The best that dsub can do is to verify that some output was created for each such parameter.

While it's allowed to specify --output and --tasks on the command line at the same time (for example if output has wildcards and each task writes a different file that matches the pattern), note that in this scenario --skip will dutifully skip all tasks if any output matching the pattern is present. In practice this means that it's generally unwise to use all three of --output, --tasks, and --skip on the command line. Instead, specify a different output for each task, in the tasks file.