slurmzy

Slurm wrapper scripts.

Install

It's pure python, with no dependencies (apart from assuming Slurm is installed, obviously). Run:

python3 -m pip install .

to install the script slurmzy

Acknowledgements

This repo contains a copy of the excellent jobinfo script from https://github.com/birc-aeh/slurm-utils/tree/master

tl;dr synopsis

Submit a job asking for 1GB RAM, called my_job, that runs the script my_script:

slurmzy run 1 my_job my_script

It will write stdout, and job stats (eg run time, memory etc) to the file my_job.o, and stderr to the file my_job.e.

Get a summary of the job, when it has finished:

slurmzy ostats my_job.o

Submit job

Run a job with:

slurmzy run <memory in GB> <name> <command>

Notes:

By default it sets a time limit of 8 hours. Change this to limit to N hours with --time N.
The default queue/partition is slurm's default, unless you have the environment variable SLURMZY_DEFAULT_PARTITION set, in which case that is used. Either way, the option --queue overrides the default.

The job will write all stdout to the file name.o, and all stderr to name.e. It will also add some job stats at the end of name.o: the output of /usr/bin/time -v, and some more stats. This is a basic attempt to reproduce what LSF does (but more grep-friendly). It also gets the output of jobinfo on the job at the end, but since the job has not yet finished, the output is of limited use (and the state is probably "RUNNING").

The options to run are:

  --norun               Do not submit job. Print the script that would be submitted
  --array_start INT     Start index of job array
  --array_end INT       End index of job array
  --array_limit INT     Max array elements allowed to run [10]
  -c INT, --cpus INT    Number of cpus [1]
  -q QUEUE_NAME, --queue QUEUE_NAME
                        Queue ('partition') to use instead of default
  -t FLOAT, --time FLOAT
                        Time limit in hours [1]
  --afterok JOB_ID[,JOB_ID2,...]
                        Only run this job after specified job(s) finish ok.
                        Provide a comma-separated list of job IDs

Job arrays

To run a job array, use the --array_start and --array_end options to specify the first and last indexes. They must both be used. The option --array_limit limits the total job elements allowed to run at any time.

To help stop driving yourself mad with escaping special characters, if you put SLURM_ARRAY_TASK_ID in the command, it will be replaced with $SLURM_ARRAY_TASK_ID (ie $ added) in the command that is run.

Toy example:

slurmzy run --array_start 1 --array_end 3 0.1 name echo array index is SLURM_ARRAY_TASK_ID

This submits a job array with 3 elements. Each job simply runs this command:

echo array index is $SLURM_ARRAY_TASK_ID

Job 1 prints array index is 1, and similarly for jobs 2 and 3.

The output files are called name.1.o, name.2.o, name.3.o, and similarly for the .e error files.

Get stats of finished jobs

Run this to get a summary of jobs from a list of output .o files:

slurmzy ostats *.o

That will summarise all .o files in your current directory. It outputs to stdout in TSV format. Example:

slurmzy ostats -f *.o | column -t
exit_code  system_time_h  wall_clock_h  max_ram  requested_ram  nodes  filename
0          0.2            0.21          1.01     1.6            node1  happy.o
137        1.0            1.1           2.01     1.0            node2  too_much_ram.o
TIMEOUT    None           0.02          None     0.1            node3  hit_time_limit.o

In that example, the first job ran OK (exit_code zero), the second hit the memory limit, and the third job hit the time limit. Except for TIMEOUT, the exit_code will be the actual exit code from the command you ran.

The RAM is in GB.

When everything runs ok, all the columns should have sensible entries. When things fail, some columns do not have information (this is a work in progess and things may improve).

Any .o files that have no run information will be ignored, and not appear in the output. This means you can safely get the stats of jobs where you know some will still be running. For example running slurmzy ostats *.o gets stats of all the finished jobs in your current directory and not report on the jobs that are in progress.

Alternatively, get an even shorter summary that shows just the exit codes of all the input *.o files. This count does include jobs that have no run information, with the exit code being No_data. Example:

slurmzy ostats --summary *.o
exit_code  count
No_data    2
0          42
137        1

Options to ostats are:

  -a, --all_columns   Output all columns
  -f, --fails         Output only failed jobs
  -s, --summary       Output summary of exit codes instead of one line per .o file
  --time_units s|m|h  Time units to report, h (hours), m (minutes), s (seconds) [h]

Known issues

RAM and cpu time are not reported by ostats if the job fails because of a timeout or because it was cancelled with scancel.

The output of jobinfo is put into the .o file. This is the last thing that is run, but the job is still running under slurm. This means the contents are not reliable, particularly the RAM and CPU stats. For this reason, we get the CPU and memory from the output of /usr/bin/time -v. However, this output is not always present if the job is terminated prematurely.

alienzj / slurmzy