flux-mini(1)

SYNOPSIS

flux mini submit [OPTIONS] [--ntasks=N] COMMAND...

flux mini bulksubmit [OPTIONS] [--ntasks=N] COMMAND...

flux mini run [OPTIONS] [--ntasks=N] COMMAND...

flux mini batch [OPTIONS] --nslots=N SCRIPT...

flux mini alloc [OPTIONS] --nslots=N [COMMAND...]

DESCRIPTION

flux-mini(1) submits jobs to run under Flux. In the case of submit or run the job consists of N copies of COMMAND launched together as a parallel job, while batch and alloc submit a script or launch a command as the initial program of a new Flux instance.

If --ntasks is unspecified, a value of N=1 is assumed. Commands that take --nslots have no default and require that --nslots or --nodes be specified.

The submit and batch commands enqueue the job and print its numerical Job ID on standard output.

The run and alloc commands do the same interactively, blocking until the job has completed.

The bulksubmit command enqueues one job each for a set of inputs read on either stdin, or given on the command line. The inputs are optionally substituted in COMMAND and/or many submission options. See more in the BULKSUBMIT section below.

For flux-mini batch, the SCRIPT given on the command line is assumed to be a file name, unless the --wrap option used, and the script file is read and submitted along with the job. If no SCRIPT is provided, then one will be read from stdin.

flux-mini alloc works similarly to batch, but instead blocks until the job has started and interactively attaches to the new Flux instance (unless the --bg option is used). By default, a new shell is spawned as the initial program of the instance, but this may be overridden by supplying COMMAND on the command line.

The intent is for the "mini" commands to remain simple with stable interfaces over time, making them suitable for use in scripts.

The available OPTIONS are detailed below.

Note

These commands target the enclosing instance. For example, within a SCRIPT submitted with flux mini batch, a flux mini command will submit a job to the batch instance, not the parent instance under which the batch instance is running. To target the parent instance, the --parent option of flux(1) should be used, e.g:

flux --parent mini batch [OPTIONS].. ARGS..

JOB PARAMETERS

These commands accept only the simplest parameters for expressing the size of the parallel program and the geometry of its task slots:

Common resource options

All subcommands take the following common resource allocation options:

-N, --nodes=N

Set the number of nodes to assign to the job. Tasks will be distributed evenly across the allocated nodes, unless the per-resource options (noted below) are used with submit, run, or bulksubmit. It is an error to request more nodes than there are tasks. If unspecified, the number of nodes will be chosen by the scheduler.

-x, --exclusive

Indicate to the scheduler that nodes should be exclusively allocated to this job. It is an error to specify this option without also using -N, --nodes. If --nodes is specified without --nslots or --ntasks, then this option will be enabled by default and the number of tasks or slots will be set to the number of requested nodes.

Per-task options

The run, submit and bulksubmit commands take two sets of mutually exclusive options to specify the size of the job request. The most common form uses the total number of tasks to run along with the amount of resources required per task to specify the resources for the entire job:

-n, --ntasks=N

Set the number of tasks to launch (default 1).

-c, --cores-per-task=N

Set the number of cores to assign to each task (default 1).

-g, --gpus-per-task=N

Set the number of GPU devices to assign to each task (default none).

Per-resource options

The second set of options allows an amount of resources to be specified with the number of tasks per core or node set on the command line. It is an error to specify any of these options when using any per-task option listed above:

--cores=N

Set the total number of cores.

--tasks-per-node=N

Set the number of tasks per node to run.

--gpus-per-node=N

With -N, --nodes, request a specific number of GPUs per node.

--tasks-per-core=N

Force a number of tasks per core. Note that this will run N tasks per allocated core. If nodes are exclusively scheduled by configuration or use of the --exclusive flag, then this option could result in many more tasks than expected. The default for this option is effectively 1, so it is useful only for oversubscribing tasks to cores for testing purposes. You probably don't want to use this option.

Batch job options

The batch and alloc commands do not launch tasks directly, and therefore job parameters are specified in terms of resource slot size and number of slots. A resource slot can be thought of as the minimal resources required for a virtual task. The default slot size is 1 core.

-n, --nslots=N

Set the number of slots requested. This parameter is required.

-c, --cores-per-slot=N

Set the number of cores to assign to each slot (default 1).

-g, --gpus-per-slot=N

Set the number of GPU devices to assign to each slot (default none).

Additional job options

The run, submit, batch, and alloc commands also take following additional job parameters:

-q, --queue=NAME

Submit a job to a specific named queue. If a queue is not specified and queues are configured, then the jobspec will be modified at ingest to specify the default queue. If queues are not configured, then this option is ignored, though flux-jobs(1) may display the queue name in its rendering of the {queue} attribute.

-t, --time-limit=MINUTES|FSD

Set a time limit for the job in either minutes or Flux standard duration (RFC 23). FSD is a floating point number with a single character units suffix ("s", "m", "h", or "d"). The default unit for the --time-limit option is minutes when no units are otherwise specified. If the time limit is unspecified, the job is subject to the system default time limit.

--job-name=NAME

Set an alternate job name for the job. If not specified, the job name will default to the command or script executed for the job.

--flags=FLAGS

Set comma separated list of job submission flags. The possible flags are waitable, novalidate, and debug. The waitable flag will allow the job to be waited on via flux job wait and similar API calls. The novalidate flag will inform flux to skip validation of a job's specification. This may be useful for high throughput ingest of a large number of jobs. Both waitable and novalidate require instance owner privileges. debug will output additional debugging into the job eventlog.

STANDARD I/O

By default, task stdout and stderr streams are redirected to the KVS, where they may be accessed with the flux job attach command.

In addition, flux-mini run processes standard I/O in real time, emitting the job's I/O to its stdout and stderr.

--input=FILENAME

Redirect stdin to the specified filename, bypassing the KVS.

--output=TEMPLATE

Specify the filename TEMPLATE for stdout redirection, bypassing the KVS. TEMPLATE may be a mustache template which supports the tags {{id}} and {{jobid}} which expand to the current jobid in the F58 encoding. If needed, an alternate encoding can be selected by using a subkey with the name of the desired encoding, e.g. {{id.dec}}. Supported encodings include f58 (the default), dec, hex, dothex, and words. For flux mini batch the default TEMPLATE is flux-{{id}}.out. To force output to KVS so it is available with flux job attach, set TEMPLATE to none or kvs.

--error=TEMPLATE

Redirect stderr to the specified filename TEMPLATE, bypassing the KVS. TEMPLATE is expanded as described above.

-l, --label-io

Add task rank prefixes to each line of output.

CONSTRAINTS

Note

Flux supports an advanced constraint specification detailed in RFC 31. However, the interface currently exported via the flux mini commands is purposefully limited.

--requires=LIST

Specify a LIST of resource property constraints for this job. LIST is a single property or comma-separated list of properties which are required for this job. The --requires option may be specified multiple times. Currently, all properties are required (logical and). If a property name starts with ^, then the job requires that property not be present on assigned resources.

DEPENDENCIES

Note

Flux supports a simple but powerful job dependency specification in jobspec. See Flux Framework RFC 26 for more detailed information about the generic dependency specification.

Dependencies may be specified on the flux mini command line using the following option

--dependency=URI

Specify a dependency of the submitted job using RFC 26 dependency URI format. The URI format is SCHEME:VALUE[?key=val[&key=val...]]. The URI will be converted into RFC 26 JSON object form and appended to the jobspec attributes.system.dependencies array. If the current Flux instance does not support dependency scheme SCHEME, then the submitted job will be rejected with an error message indicating this fact.

The --dependency option may be specified multiple times. Each use appends a new dependency object to the attributes.system.dependencies array.

The following dependency schemes are built-in:

Note

The after* dependency schemes listed below all require that the target JOBID be currently active or in the job manager's inactive job cache. If a target JOBID has been purged by the time the dependent job has been submitted, then the submission will be rejected with an error that the target job cannot be found.

after:JOBID

This dependency is satisfied after JOBID starts.

afterany:JOBID

This dependency is satisfied after JOBID enters the INACTIVE state, regardless of the result

afterok:JOBID

This dependency is satisfied after JOBID enters the INACTIVE state with a successful result.

afternotok:JOBID

This dependency is satisfied after JOBID enters the INACTIVE state with an unsuccessful result.

begin-time:TIMESTAMP

This dependency is satisfied after TIMESTAMP, which is specified in floating point seconds since the UNIX epoch. See the flux-mini --begin-time option below for a more user-friendly interface to the begin-time dependency.

In any of the above after* cases, if it is determined that the dependency cannot be satisfied (e.g. a job fails due to an exception with afterok), then a fatal exception of type=dependency is raised on the current job.

ENVIRONMENT

By default, flux-mini duplicates the current environment when submitting jobs. However, a set of environment manipulation options are provided to give fine control over the requested environment submitted with the job.

--env=RULE

Control how environment variables are exported with RULE. See ENV RULE SYNTAX section below for more information. Rules are applied in the order in which they are used on the command line. This option may be specified multiple times.

--env-remove=PATTERN

Remove all environment variables matching PATTERN from the current generated environment. If PATTERN starts with a / character, then it is considered a regex(7), otherwise PATTERN is treated as a shell glob(7). This option is equivalent to --env=-PATTERN and may be used multiple times.

--env-file=FILE

Read a set of environment RULES from a FILE. This option is equivalent to --env=^FILE and may be used multiple times.

ENV RULES

The --env* options of flux-mini allow control of the environment exported to jobs via a set of RULE expressions. The currently supported rules are

  • If a rule begins with -, then the rest of the rule is a pattern which removes matching environment variables. If the pattern starts with /, it is a regex(7), optionally ending with /, otherwise the pattern is considered a shell glob(7) expression.

    Examples:

    -* or -/.*/ filter all environment variables creating an empty environment.

  • If a rule begins with ^ then the rest of the rule is a filename from which to read more rules, one per line. The ~ character is expanded to the user's home directory.

    Examples:

    ~/envfile reads rules from file $HOME/envfile

  • If a rule is of the form VAR=VAL, the variable VAR is set to VAL. Before being set, however, VAL will undergo simple variable substitution using the Python string.Template class. This simple substitution supports the following syntax:

    • $$ is an escape; it is replaced with $

    • $var will substitute var from the current environment, falling back to the process environment. An error will be thrown if environment variable var is not set.

    • ${var} is equivalent to $var

    • Advanced parameter substitution is not allowed, e.g. ${var:-foo} will raise an error.

    Examples:

    PATH=/bin, PATH=$PATH:/bin, FOO=${BAR}something

  • Otherwise, the rule is considered a pattern from which to match variables from the process environment if they do not exist in the generated environment. E.g. PATH will export PATH from the current environment (if it has not already been set in the generated environment), and OMP* would copy all environment variables that start with OMP and are not already set in the generated environment. It is important to note that if the pattern does not match any variables, then the rule is a no-op, i.e. an error is not generated.

    Examples:

    PATH, FLUX_*_PATH, /^OMP.*/

Since flux-mini always starts with a copy of the current environment, the default implicit rule is * (or --env=*). To start with an empty environment instead, the -* rule or --env-remove=* option should be used. For example, the following will only export the current PATH to a job:

flux mini run --env-remove=* --env=PATH ...

Since variables can be expanded from the currently built environment, and --env options are applied in the order they are used, variables can be composed on the command line by multiple invocations of --env, e.g.:

flux mini run --env-remove=* \
              --env=PATH=/bin --env='PATH=$PATH:/usr/bin' ...

Note that care must be taken to quote arguments so that $PATH is not expanded by the shell.

This works particularly well when specifying rules in a file:

-*
OMP*
FOO=bar
BAR=${FOO}/baz

The above file would first clear the environment, then copy all variables starting with OMP from the current environment, set FOO=bar, and then set BAR=bar/baz.

PROCESS RESOURCE LIMITS

By default flux mini propagates some common resource limits (as described in getrlimit(2)) to the job by setting the rlimit job shell option in jobspec. The set of resource limits propagated can be controlled via the --rlimit=RULE option:

--rlimit=RULE

Control how process resource limits are propagated with RULE. Rules are applied in the order in which they are used on the command line. This option may be used multiple times.

The --rlimit rules work similar to the --env option rules:

  • If a rule begins with -, then the rest of the rule is a name or glob(7) pattern which removes matching resource limits from the set to propagate.

    Example:

    -* disables propagation of all resource limits.

  • If a rule is of the form LIMIT=VALUE then LIMIT is explicitly set to VALUE. If VALUE is unlimited, infinity or inf, then the value is set to RLIM_INFINITY (no limit).

    Example:

    nofile=1024 overrides the current RLIMIT_NOFILE limit to 1024.

  • Otherwise, RULE is considered a pattern from which to match resource limits and propagate the current limit to the job, e.g.

    --rlimit=memlock

    will propagate RLIMIT_MEMLOCK (which is not in the list of limits that are propagated by default).

flux-mini starts with a default list of resource limits to propagate, then applies all rules specified via --rlimit on the command line. Therefore, to propagate only one limit, -* should first be used to start with an empty set, e.g. --rlimit=-*,core will only propagate the core resource limit.

The set of resource limits propagated by default includes all those except memlock, ofile, msgqueue, nice, rtprio, rttime, and sigpending. To propagate all possible resource limits, use --rlimit=*.

EXIT STATUS

The job exit status, normally the largest task exit status, is stored in the KVS. If one or more tasks are terminated with a signal, the job exit status is 128+signo.

The flux-job attach command exits with the job exit status.

In addition, flux-mini run runs until the job completes and exits with the job exit status.

OTHER OPTIONS

--cwd=DIRECTORY

Set job working directory.

--urgency=N

Specify job urgency, which affects queue order. Numerically higher urgency jobs are considered by the scheduler first. Guests may submit jobs with urgency in the range of 0 to 16, while instance owners may submit jobs with urgency in the range of 0 to 31 (default 16). In addition to numerical values, the special names hold (0), default (16), and expedite (31) are also accepted.

-v, --verbose

(run,alloc,submit,bulksubmit) Increase verbosity on stderr. For example, currently flux mini run -v displays jobid, -vv displays job events, and -vvv displays exec events. flux mini alloc -v forces the command to print the submitted jobid on stderr. The specific output may change in the future.

-o, --setopt=KEY[=VAL]

Set shell option. Keys may include periods to denote hierarchy. VAL is optional and may be valid JSON (bare values, objects, or arrays), otherwise VAL is interpreted as a string. If VAL is not set, then the default value is 1. See SHELL OPTIONS below.

--setattr=KEY[=VAL]

Set jobspec attribute. Keys may include periods to denote hierarchy. If KEY does not begin with system., user., or ., then system. is assumed. VAL is optional and may be valid JSON (bare values, objects, or arrays), otherwise VAL is interpreted as a string. If VAL is not set, then the default value is 1. If KEY starts with a ^ character, then VAL is interpreted as a file, which must be valid JSON, to use as the attribute value.

--begin-time=DATETIME

Convenience option for setting a begin-time dependency for a job. The job is guaranteed to start after the specified date and time. If DATETIME begins with a + character, then the remainder is considered to be an offset in Flux standard duration (RFC 23), otherwise, any datetime expression accepted by the Python parsedatetime module is accepted, e.g. 2021-06-21 8am, in an hour, tomorrow morning, etc.

--taskmap=SCHEME[:VALUE]

Choose an alternate method for mapping job task IDs to nodes of the job. The job shell maps tasks using a "block" distribution scheme by default (consecutive tasks share nodes) This option allows the activation of alternate schemes by name, including an optional VALUE. Supported schemes which are built in to the job shell include

cyclic[:N]

Tasks are distributed over consecutive nodes with a stride of N (where N=1 by default).

manual:TASKMAP

An explicit RFC 34 taskmap is provided and used to manually map task ids to nodes. The provided TASKMAP must match the total number of tasks in the job and the number of tasks per node assigned by the job shell, so this option is not useful unless the total number of nodes and tasks per node are known at job submission time.

However, shell plugins may provide other task mapping schemes, so check the current job shell configuration for a full list of supported taskmap schemes.

--dry-run

Don't actually submit job. Just emit jobspec on stdout and exit for run, submit, alloc, and batch. For bulksubmit, emit a line of output including relevant options for each job which would have been submitted,

--debug

Enable job debug events, primarily for debugging Flux itself. The specific effects of this option may change in the future.

--bg

(alloc only) Do not interactively attach to the instance. Instead, print jobid on stdout once the instance is ready to accept jobs. The instance will run indefinitely until a time limit is reached, the job is canceled, or it is shutdown with flux shutdown JOBID (preferred). If a COMMAND is given then the job will run until COMMAND completes. Note that flux job attach JOBID cannot be used to interactively attach to the job (though it will print any errors or output).

-B, --broker-opts=OPT

(batch only) For batch jobs, pass specified options to the Flux brokers of the new instance. This option may be specified multiple times.

--wrap

(batch only) The --wrap option wraps the specified COMMAND and ARGS in a shell script, by prefixing with #!/bin/sh. If no COMMAND is present, then a SCRIPT is read on stdin and wrapped in a /bin/sh script.

--cc=IDSET

(submit,bulksubmit) Replicate the job for each id in IDSET. FLUX_JOB_CC=id will be set in the environment of each submitted job to allow the job to alter its execution based on the submission index. (e.g. for reading from a different input file). When using --cc, the substitution string {cc} may be used in options and commands and will be replaced by the current id.

--bcc=IDSET

(submit,bulksubmit) Identical to --cc, but do not set FLUX_JOB_CC in each job. All jobs will be identical copies. As with --cc, {cc} in option arguments and commands will be replaced with the current id.

--quiet

(submit,bulksubmit) Suppress logging of jobids to stdout.

--log=FILE

(submit,bulksubmit) Log flux-mini output and stderr to FILE instead of the terminal. If a replacement (e.g. {} or {cc}) appears in FILE, then one or more output files may be opened. For example, to save all submitted jobids into separate files, use:

flux mini submit --cc=1-4 --log=job{cc}.id hostname
--log-stderr=FILE

(submit,bulksubmit) Separate stderr into FILE instead of sending it to the terminal or a FILE specified by --log.

--wait

(submit,bulksubmit) Wait on completion of all jobs before exiting. This is equivalent to --wait-event=clean.

--wait-event=NAME

(run,submit,bulksubmit) Wait until job or jobs have received event NAME before exiting. E.g. to submit a job and block until the job begins running, use --wait-event=start. (submit,bulksubmit only) If NAME begins with exec., then wait for an event in the exec eventlog, e.g. exec.shell.init. For flux mini run the argument to this option when used is passed directly to flux job attach.

--watch

(submit,bulksubmit) Display output from all jobs. Implies --wait.

--progress

(submit,bulksubmit) With --wait, display a progress bar showing the progress of job completion. Without --wait, the progress bar will show progress of job submission.

--jps

(submit,bulksubmit) With --progress, display throughput statistics (jobs/s) in the progress bar.

--define=NAME=CODE

(bulksubmit) Define a named method that will be made available as an attribute during command and option replacement. The string being processed is available as x. For example:

$ seq 1 8 | flux mini bulksubmit --define=pow="2**int(x)" -n {.pow} ...
--shuffle

(bulksubmit) Shuffle the list of commands before submission.

--sep=STRING

(bulksubmit) Change the separator for file input. The default is to separate files (including stdin) by newline. To separate by consecutive whitespace, specify --sep=none.

BULKSUBMIT

The bulksubmit utility allows rapid bulk submission of jobs using an interface similar to GNU parallel or xargs. The command takes inputs on stdin or the command line (separated by :::), and submits the supplied command template and options as one job per input combination.

The replacement is done using Python's string.format(), which is supplied a list of inputs on each iteration. Therefore, in the common case of a single input list, {} will work as the substitution string, e.g.:

$ seq 1 4 | flux mini bulksubmit echo {}
flux-mini: submit echo 1
flux-mini: submit echo 2
flux-mini: submit echo 3
flux-mini: submit echo 4

With --dry-run bulksubmit will print the args and command which would have been submitted, but will not perform any job submission.

The bulksubmit command can also take input lists on the command line. The inputs are separated from each other and the command with the special delimiter ::::

$ flux mini bulksubmit echo {} ::: 1 2 3 4
flux-mini: submit echo 1
flux-mini: submit echo 2
flux-mini: submit echo 3
flux-mini: submit echo 4

Multiple inputs are combined, in which case each input is passed as a positional parameter to the underlying format(), so should be accessed by index:

$ flux mini bulksubmit --dry-run echo {1} {0} ::: 1 2 ::: 3 4
flux-mini: submit echo 3 1
flux-mini: submit echo 4 1
flux-mini: submit echo 3 2
flux-mini: submit echo 4 2

If the generation of all combinations of an input list with other inputs is not desired, the special input delimited :::+ may be used to "link" the input, so that only one argument from this source will be used per other input, e.g.:

$ flux mini bulksubmit --dry-run echo {0} {1} ::: 1 2 :::+ 3 4
flux-mini: submit 1 3
flux-mini: submit 2 4

The linked input will be cycled through if it is shorter than other inputs.

An input list can be read from a file with :::::

$ seq 0 3 >inputs
$ flux mini bulksubmit --dry-run :::: inputs
flux-mini: submit 0
flux-mini: submit 1
flux-mini: submit 2
flux-mini: submit 3

If the filename is - then stdin will be used. This is useful for including stdin when reading other inputs.

The delimiter ::::+ indicates that the next file is to be linked to the inputs instead of combined with them, as with :::+.

There are several predefined attributes for input substitution. These include:

  • {.%} returns the input string with any extension removed.

  • {./} returns the basename of the input string.

  • {./%} returns the basename of the input string with any extension removed.

  • {.//} returns the dirname of the input string

  • {seq} returns the input sequence number (0 origin)

  • {seq1} returns the input sequence number (1 origin)

  • {cc} returns the current id from use of --cc or --bcc. Note that replacement of {cc} is done in a second pass, since the --cc option argument may itself be replaced in the first substitution pass. If --cc/bcc were not used, then {cc} is replaced with an empty string. This is the only substitution supported with flux-mini submit.

Note that besides {seq}, {seq1}, and {cc} these attributes can also take the input index, e.g. {0.%} or {1.//}, when multiple inputs are used.

Additional attributes may be defined with the --define option, e.g.:

$ flux mini bulksubmit --dry-run --define=p2='2**int(x)' -n {.p2} hostname \
   ::: $(seq 0 4)
flux-mini: submit -n1 hostname
flux-mini: submit -n2 hostname
flux-mini: submit -n4 hostname
flux-mini: submit -n8 hostname
flux-mini: submit -n16 hostname

The input string being indexed is passed to defined attributes via the local x as seen above.

SHELL OPTIONS

These options are provided by built-in shell plugins that may be overridden in some cases:

mpi=spectrum

Load the MPI personality plugin for IBM Spectrum MPI. All other MPI plugins are loaded by default.

cpu-affinity=per-task

Tasks are distributed across the assigned resources.

cpu-affinity=off

Disable task affinity plugin.

gpu-affinity=per-task

GPU devices are distributed evenly among local tasks. Otherwise, GPU device affinity is to the job.

gpu-affinity=off

Disable GPU affinity for this job.

verbose

Increase verbosity of the job shell log.

nosetpgrp

Normally the job shell runs each task in its own process group to facilitate delivering signals to tasks which may call fork(2). With this option, the shell avoids calling setpgrp(2), and each task will run in the process group of the shell. This will cause signals to be delivered only to direct children of the shell.

pmi=off

Disable the process management interface (PMI-1) which is required for bootstrapping most parallel program environments. See flux-shell(1) for more pmi options.

stage-in

Copy files previously mapped with flux-filemap(1) to $FLUX_JOB_TMPDIR. See flux-shell(1) for more stage-in options.

RESOURCES

Flux: http://flux-framework.org