A python script for tabular display of slurm efficiency information
Whether a sys admin or cluster user, knowing how well you are estimating job
resources can help streamline job scheduling and maximize your priority. If you
have ever tried to use sacct you probably had some trouble interpreting the
output. While seff or jobstats can provide detailed summaries, they don't
scale easily to array jobs or offer a way to see all the jobs from a single
user. reportseff aims to fill this role. Read more about the motivation
for reportseff.
If you are running more than one slurm job at a time, you should try
reportseff. Users of HPC systems can get an idea how well they estimate
resource usage. By tuning these values, you can get scheduled earlier and not
be penalized for unused allocations. Since reportseff can parse job ids from
slurm output files, it simplifies the task of identifying which jobs have
failed and why. Sys admins can pipe reportseff output to identify users with
poor utilization or produce summaries at the end of a billing cycle.
reportseff is a wrapper around sacct that provides more complex option
parsing, simpler options, and cleaner, colored outputs. All querying is
performed in a single call to sacct and should have similar performance.
Multi-node and GPU utilization is acquired from information contained in the
AdminComment field, as generated by jobstats.
reportseff runs on python >= 3.6.
The only external dependency is click (>= 6.7).
Calling
pip install --user reportseff
# OR
pipx install reportseffwill create command line bindings and install click.
Try reportseff -u $USER or just reportseff in a directory with some slurm
outputs. You may be surprised by your results!
Calling reportseff with a single jobid will provide equivalent information to
seff for that job. reportseff 24371789 and reportseff map_fastq_24371789
produce the following output:
JobID State Elapsed CPUEff MemEff
24371789 COMPLETED 03:08:03 71.2% 45.7%Providing either the raw job id or the array job id will get efficiency
information for a single element of the array job. reportseff 24220929_421
and reportseff 24221219 generate:
JobID State Elapsed CPUEff MemEff
24220929_421 COMPLETED 00:09:34 99.0% 34.6%If the base job id of an array is provided, all elements of the array will
be added to the output. reportseff 24220929
JobID State Elapsed CPUEff MemEff
24220929_1 COMPLETED 00:10:43 99.2% 33.4%
24220929_11 COMPLETED 00:10:10 99.2% 37.5%
24220929_21 COMPLETED 00:09:25 98.8% 36.1%
24220929_31 COMPLETED 00:09:19 98.9% 33.3%
24220929_41 COMPLETED 00:09:23 98.9% 33.3%
24220929_51 COMPLETED 00:08:02 98.5% 36.3%
...
24220929_951 COMPLETED 00:25:12 99.5% 33.5%
24220929_961 COMPLETED 00:39:26 99.7% 34.1%
24220929_971 COMPLETED 00:24:11 99.5% 34.2%
24220929_981 COMPLETED 00:24:50 99.5% 44.3%
24220929_991 COMPLETED 00:13:05 98.7% 33.7%Because slurm output files can act as job id inputs, the following can get all seff information for a given job name:
slurm_out ❯❯❯ reportseff split_ubam_24*
JobID State Elapsed CPUEff MemEff
split_ubam_24342816 COMPLETED 23:30:32 99.9% 4.5%
split_ubam_24342914 COMPLETED 22:40:51 99.9% 4.6%
split_ubam_24393599 COMPLETED 23:43:36 99.4% 4.4%
split_ubam_24393655 COMPLETED 21:36:58 99.3% 4.5%
split_ubam_24418960 RUNNING 02:53:11 --- ---
split_ubam_24419972 RUNNING 01:26:26 --- ---Without arguments, reportseff will try to find slurm output files in the
current directory. Combine with watch to monitor job progress:
watch -cn 300 reportseff --color --modified-sort
JobID State Elapsed CPUEff MemEff
split_ubam_24418960 RUNNING 02:56:14 --- ---
fastq_to_ubam_24419971 RUNNING 01:29:29 --- ---
split_ubam_24419972 RUNNING 01:29:29 --- ---
fastq_to_ubam_24393600 COMPLETED 1-02:00:47 58.3% 41.1%
map_fastq_24419330 RUNNING 02:14:53 --- ---
map_fastq_24419323 RUNNING 02:15:24 --- ---
map_fastq_24419324 RUNNING 02:15:24 --- ---
map_fastq_24419322 RUNNING 02:15:24 --- ---
mark_adapters_24418437 COMPLETED 01:29:23 99.8% 48.2%
mark_adapters_24418436 COMPLETED 01:29:03 99.9% 47.4%One useful application of reportseff is filtering a directory of slurm output
files based on the state or time since running. Additionally, if only the
jobid is specified as a format output, the filenames will be returned in a
pipe-friendly manner:
old_runs ❯❯❯ reportseff --since d=4 --state Timeout
JobID State Elapsed CPUEff MemEff
call_variants_31550458 TIMEOUT 20:05:17 99.5% 0.0%
call_variants_31550474 TIMEOUT 20:05:17 99.6% 0.0%
call_variants_31550500 TIMEOUT 20:05:08 99.7% 0.0%
old_runs ❯❯❯ reportseff --since d=4 --state Timeout --format jobid
call_variants_31550458
call_variants_31550474
call_variants_31550500To find all lines with output: in jobs which have timed out or failed
in the last 4 days:
reportseff --since 'd=4' --state TO,F --format jobid | xargs grep output:Jobs can be passed as arguments in the following ways:
- Job ID such as 1234567. If the id is part of an array job, only the element for that ID will be displayed. If the id is the base part of an array job, all elements in the array will be displayed.
- Array Job ID such as 1234567_89. Will display only the element specified.
- Slurm output file. Format must be BASE_%A_%a. BASE is optional as is a '.out' suffix. Unix glob expansions can also be used to filter which jobs are displayed.
- From current directory. If no argument is supplied,
reportseffwill attempt to find slurm output files in the current directory as described above. If a user is provided, insteadreportseffwill show recent jobs for that user. If onlysinceis set, all recent jobs for all users will be shown (if allowed). - Supplying a directory as a single argument will override the current directory to check for slurm outputs.
--color/--no-color: Force color output or not. By default, will force color output. With the no-color flag, click will strip color codes for everything besides stdout.--modified-sort: Instead of sorting by filename/jobid, sort by last modification time of the slurm output file.--debug: Write sacct result to stderr.--user/-u: Ignore job arguments and instead query sacct with provided user. Returns all jobs from the last week.--state/-s: Output only jobs with states matching one of the provided options. Accepts comma separated values of job codes (e.g. 'R') or full names (e.g. RUNNING). Case insensitive.--not-state/-S: Output only jobs with states not matching any of the provided options. Accepts comma separated values of job codes (e.g. 'R') or full names (e.g. RUNNING). Case insensitive.--format: Provide a comma separated list of columns to produce. Prefixing the argument with+adds the specified values to the defaults. Values can be any valid column name to sacct and the custom efficiency values: TimeEff, cpuEff, MemEff. Can also optionally set alignment (<, ^, >) and maximum width. Default is center-aligned with a width of the maximum column entry. For example,--format 'jobid%>,state%10,memeff%<5'produces 3 columns with:- JobId aligned right, width set automatically
- State with width 10 (center aligned by default)
- MemEff aligned left, width 5
--slurm-format: The filename pattern passed to sbatch during job submission. Overrides the default regex for job id parsing from filenames. E.g. to match filenames like123456.outset--slurm-format %j.out.--since: Limit results to those occurring after the specified time. Accepts sacct formats and a comma separated list of key/value pairs. To get jobs in the last hour and a half, can passh=1,m=30. ---until: Limit results to those occurring before the specified time. Accepts sacct formats and a comma separated list of key/value pairs. Useful in combination with the 'since' option to query a specific range.--partition: Limit results to a specific partition.--node/-n: Display information for multi-node jobs; requires additional sacct fields from jobstats.--node-and-gpu/-g: Display information for multi-node jobs and GPU information; requires additional sacct fields from jobstats.--parsable/-p: Ignore formatting and output as a|delimited table. Useful for piping into more complex analyses.
reportseff is actively maintained but currently feature complete. If there
is a function missing, please open an issue to discuss its merit!
Bug reports, pull requests, and any feedback are welcome! Prior to submitting a pull request, be sure any new features have been tested and all unit tests are passing. In the cloned repo with poetry installed:
poetry install
poetry run pytest
poetry run pre-commit install
noxpip is the package installer for python. If you get an error that pip isn't found, look for a python/anaconda/conda module. pipx ensures that each application is installed in an isolated environment. This resolves issues of dependency versions and allows applications to be run from any environment.
Click should determine if the output supports color display and react automatically
in a way you expect. Check that your terminal is setup to display colors and
that your pager (probably less) will display color by default. Some commands,
e.g. watch aren't handled properly even when invoked to support color. Here
are some useful settings for your .bashrc:
# have less display colors by default. Will fix `reportseff` not showing colors
export LESS="-R"
# for watch aliases, include the `--color` option
watch -cn 300 reportseff --color --modified-sort
# ^ ^^^^^^^
You can always for display of color (or suppress it) with the --color/--no-color
options
Python will report that the consumer of process output has closed the stream (i.e. the pipe) while still attempting to write. Newer versions of click should suppress the warning output, but it seems to not always work. Besides some extra printing on stderr, the output is not affected.
Because sacct doesn't currently record this information, reportseff
retrieves it from a custom field from jobstats, developed at Princeton
University. If you are outside a Research Computing cluster, that information
will likely be absent. Node-level reporting is only shown for jobs which use
multiple nodes or GPUs. If you need a list of where jobs were run, you can add
--format +NodeList.
The code for calling sacct and parsing the returning information was taken from Slurmee.
Style and tooling from hypermodern python
Code review provided from a repo-review which vastly improved this readme.