Flux Administrator’s Guide

The Flux Administrator’s Guide documents relevant information for installation, configuration, and management of Flux as the native resource manager on a cluster.


Flux is still beta software and many of the interfaces documented in this guide may change with regularity.

This document is in DRAFT form and currently applies to flux-core version 0.45.0.


0.45.0 limitation: the flux system instance is primarily tested on a 128 node cluster.

Overview and Background

Flux Architecture

A Flux instance consists of one or more Flux brokers communicating over a tree-based overlay network. Most of Flux’s distributed systems and services that aren’t directly associated with a running job are embedded in the flux-broker(1) executable or its dynamically loaded plugins.

Flux may be used in single-user mode, where a Flux instance is launched as a parallel job, and the instance owner (the user that submitted the parallel job) has control of, and exclusive access to, the Flux instance and its assigned resources. On a system running Flux natively, batch jobs and allocations are examples of single user Flux instances.

When Flux is deployed as the system instance, or native resource manager on a cluster, its brokers still run with the credentials of a non-privileged system user, typically flux. However, to support multiple users and act as a long running service, it must be configured to behave differently:

  • The Flux broker is started directly by systemd on each node instead of being launched as a process in a parallel job.

  • The systemd unit file passes arguments to the broker that tell it to use system paths for various files, and to ingest TOML files from a system configuration directory.

  • A single security certificate is used for the entire cluster instead of each broker generating one on the fly and exchanging public keys with PMI.

  • The Flux overlay network endpoints are statically configured from files instead of being generated on on the fly and exchanged via PMI.

  • The instance owner is a system account that does not correspond to an actual user.

  • Users other than the instance owner (guests) are permitted to connect to the Flux broker, and are granted limited access to Flux services.

  • Users connect to the Flux broker’s AF_UNIX socket via a well known system URI if FLUX_URI is not set in the environment.

  • Job processes (including the Flux job shell) are launched as the submitting user with the assistance of a setuid root helper on each node called the IMP.

  • Job requests are signed with MUNGE, and this signature is verified by the IMP.

  • The content of the Flux KVS, containing system state such as the set of drained nodes and the job queue, is preserved across a full Flux restart.

  • The system instance functions with some nodes offline.

  • The system instance has no initial program.

The same Flux executables are used in both single user and system modes, with operation differentiated only by configuration.

Flux system instance architecture

Fox prevents Frog from submitting jobs on a cluster with Flux as the system resource manager.

Software Components

Flux was conceived as a resource manager toolkit rather than a monolithic project, with the idea to make components like the scheduler replaceable. In addition, several parts of flux can be extended with plugins. At this time the primary component types are

broker modules

Each broker module runs in its own thread as part of the broker executable, communicating with other components using messages. Broker modules are dynamically loadable with the flux-module(1) command. Core services like the KVS, job manager, and scheduler are implemented using broker modules.

jobtap plugins

The job manager orchestrates a job’s life cycle. Jobtap plugins extend the job manager, arranging for callbacks at different points in the job life cycle. Jobtap plugins may be dynamically loaded with the flux-jobtap(1) command. An example of a jobtap plugin is the Flux accounting multi-factor priority plugin, which updates a job’s priority value when it enters the PRIORITY state.

shell plugins

When a job is started, the flux-shell(1) is the process parent of job tasks on each node. Shell plugins extend the job environment and can be configured on a per-job basis using the --setopt option of the flux-mini(1) commands. affinity, pmi, and pty are examples of Flux shell plugins.


Flux commands open a connection to a particular Flux instance by specifying a URI. The scheme portion of the URI may refer to a native connection method such as local or ssh. Native connection methods are implemented as plugins called connectors. See flux_open(3).

URI resolver plugins

Other URI schemes must be resolved to a native form before they can be used. Resolvers for new schemes may be added as plugins. For example, the lsf resolver plugin enables LSF users to connect to Flux instances running as LSF jobs by specifying a lsf:JOBID URI. See flux-uri(1).

validator plugins

Jobs may be rejected at ingest if their jobspec fails one of a set of configured validator plugins. The basic validator ensures the jobspec conforms to the jobspec specification. The feasibility plugin rejects job that the scheduler determines would be unable to run given the instance’s resource set. The require-instance plugin rejects jobs that do not run in a new Flux instance. See flux-config-ingest(5).

frobnicator plugins

The frobnicator allows a set of configured plugins to modify jobspec at submission time. For example the defaults plugin sets configured default values for jobspec attributes such as duration and queue. See flux-config-ingest(5).

Independently developed Flux components are generally packaged and versioned separately. Each package may provide one or more of the above components as well as man pages and flux(1) subcommands. At this stage of Flux development, it is good practice to combine only contemporaneously released components as the interfaces are not stable yet.

File Formats and Data Types

Since some parts of Flux are developed independently, some effort has been made to standardize file formats and data types to ensure components work together and provide a consistent user experience. System administrators may find it useful to be aware of some of them.


A compact way of representing an ordered list of hostnames, compatible with legacy tools in use at LLNL and defined by RFC 29.


A compact way of representing an unordered set of integers, defined by RFC 22.


Tom’s Oblivious Minimal Language is the file format used in Flux configuration files.


Javascript Object Notation is used throughout Flux in messages and other file formats.


An ordered log of timestamped events, stored in the Flux KVS and defined by RFC 18. Eventlogs are used to record job events, capture standard I/O streams, and record resource status changes.


Flux Standard Duration, a string format used to represent a length of time, defined by RFC 23.


A job request (JSON or YAML), defined by RFC 25 and RFC 14.


A resource set (JSON), defined by RFC 20.


Flux Locally Unique ID, used for Flux job IDs, defined by RFC 19.


The Flux brokers that make up a system instance are started on each node by systemd. The brokers run as an unprivileged system user, typically flux. This user is termed the instance owner. The instance owner has complete control of the Flux instance.

A tree-based overlay network is established among brokers, rooted at the management node. This network is secured and encrypted using the ZeroMQ CURVE mechanism. This requires a single CURVE certificate to be generated and installed on all nodes, typically /etc/flux/system/curve.cert, before Flux begins operation. The certificate must be readable by the instance owner but should be carefully protected from access by other users since disclosure could allow overlay network security to be bypassed.

On each node, users and tools may connect to the local system instance broker via a UNIX domain socket at a well known location, usually /run/flux/local. Users are authenticated on this socket using the SO_PEERCRED socket option. Once connected, a user may inject messages into the overlay network. Messages are stamped by the broker at ingress with the user’s authenticated userid, and a role mask that identifies any special capabilities granted to the user. Messages that are sent by the flux user are stamped with the instance owner role, while other users, or guests, are stamped with a role that grants minimal access. Note that the root user is considered a guest user with no special privilege in Flux, but sites can choose to grant root the owner role by configuration if desired. See flux-config-security(5).

Messages are used for remote procedure calls. A Flux service may allow or deny an RPC request depending on its message rolemask or userid. For example, only the instance owner can drain a node because the drain service only allows drain request messages that have the owner role. Similarly, any job can be canceled by a cancel request message with the owner role, but in addition, jobs can be canceled by guests whose message userid matches the target job userid.

A Flux job is launched when brokers launch one flux-shell(1) per node with the credentials of the user that submitted the job. When that is a guest user, Flux employs a setuid helper called the flux-imp(8) to launch the shells with the guest credentials. The shells in turn launch one or more user processes that compose the parallel job.

The IMP is restricted by configuration to only allow the flux user to run it, and to only launch the system Flux job shell executable. In addition, job requests are signed by the submitting user with MUNGE, and the IMP verifies this signature before starting the shells. The current working directory of the job, the environment, and the executable command line are examples of job request data protected by the MUNGE signature.

When Flux starts a batch job or allocation, it starts an independent, single-user Flux instance with brokers running as the submitting user. The new instance owner has complete control over this Flux instance, which cannot use the IMP to launch jobs as guests, and does not permit guests to connect to its UNIX domain sockets. Its overlay network is also secured with the ZeroMQ CURVE mechanism, but instead of starting with a shared certificate read from disk, each broker generates a certificate in memory on the fly, then exchanges public keys and socket endpoints with peer brokers using the PMI service offered by the Flux shells of the enclosing instance. In other words, the single-user Flux instance bootstraps like an MPI parallel program.

See also: RFC 12, RFC 15.


System Prerequisites

MUNGE is used to sign job requests submitted to Flux, so the MUNGE daemon should be installed on all nodes running Flux with the same MUNGE key used across the cluster.

Flux assumes a shared UID namespace across the cluster.

A system user named flux is required. This user need not have a valid home directory or shell.

Flux uses hwloc to verify that configured resources are present on nodes. Ensure that the system installed version includes any plugins needed for the hardware, especially GPUs.

A Word about Core Dumps

It is helpful to enable core dumps from the system instance flux-broker (especially rank 0) so that useful bug reports can be filed should the broker crash. Usually systemd-coredump(8) handles this, which makes core files and stack traces accessible with coredumpctl(1).

Some sites choose instead to configure the kernel.core_pattern sysctl(8) parameter to a relative file path, which directs core files to the program’s current working directory. Please note that the system instance broker runs as the flux user with a working directory of / and thus would not have write permission on its current working directory.


If you do observe a flux-broker crash, please open a github issue at https://github.com/flux-framework/flux-core/issues and include the Flux version, relevant log messages from journalctl -u flux, and a stack trace, if available.

Installing Software Packages

The following Flux framework packages are needed for a Flux system instance and should be installed from your Linux distribution package manager.


APIs for job signing, and the IMP, a privileged program for starting processes as multiple users. Install on all nodes (required). If building flux-security from source, be sure to configure --with-pam to include Pluggable Authentication Modules (PAM) support.


All of the core components of Flux, including the Flux broker. flux-core is functional on its own, but cannot run jobs as multiple users, has a simple FIFO scheduler, and does not implement accounting-based job prioritization. If building flux-core from source, be sure to configure with --with-flux-security. Install on all nodes (required).


The Fluxion graph-based scheduler.

flux-accounting (optional)

Management of limits for individual users/projects, banks, and prioritization based on fair-share accounting. For more information on how to configure run flux-accounting, please refer to Flux Accounting Guide.

flux-pam (optional)

A PAM module that can enable users to login to compute nodes that are running their jobs.


Flux packages are currently maintained only for the TOSS Red Hat Enterprise Linux based Linux distribution, which is not publicly distributed. Open an issue in flux-core if you would like to become a maintainer of Flux packages for another Linux distribution so we can share packaging tips and avoid duplicating effort.


Much of Flux configuration occurs via TOML configuration files found in a hierarchy under /etc/flux. There are three separate TOML configuration spaces: one for flux-security, one for the IMP (an independent component of flux-security), and one for Flux running as the system instance. Each configuration space has a separate directory, from which all files matching the glob *.toml are read. System administrators have the option of using one file for each configuration space, or breaking up each configuration space into multiple files. In the examples below, one file per configuration space is used.

For more information on the three configuration spaces, please refer to flux-config(5), flux-config-security(5), and flux-config-security-imp(5).

Configuring flux-security

When Flux is built to support multi-user workloads, job requests are signed using a library provided by the flux-security project. This library reads a static configuration from /etc/flux/security/conf.d/*.toml. Note that for security, these files and their parent directory should be owned by root without write access to other users, so adjust permissions accordingly.

Example file installed path: /etc/flux/security/conf.d/security.toml

# Job requests should be valid for 2 weeks
# Use munge as the job request signing mechanism
max-ttl = 1209600  # 2 weeks
default-type = "munge"
allowed-types = [ "munge" ]

See also: flux-config-security-sign(5).

Configuring the IMP

The Independent Minister of Privilege (IMP) is the only program that runs as root, by way of the setuid mode bit. To enhance security, it has a private configuration space in /etc/flux/imp/conf.d/*.toml. Note that the IMP will verify that files in this path and their parent directories are owned by root without write access from other users, so adjust permissions and ownership accordingly.

Example file installed path: /etc/flux/imp/conf.d/imp.toml

# Only allow access to the IMP exec method by the 'flux' user.
# Only allow the installed version of flux-shell(1) to be executed.
allowed-users = [ "flux" ]
allowed-shells = [ "/usr/libexec/flux/flux-shell" ]

# Enable the "flux" PAM stack (requires PAM configuration file)
pam-support = true

See also: flux-config-security-imp(5).

Configuring the Flux PAM Stack

If PAM support is enabled in the IMP config, the flux PAM stack must exist and have at least one auth and one session module.

Example file installed path: /etc/pam.d/flux

auth    required pam_localuser.so
session required pam_limits.so

The pam_limits.so module is useful for setting default job resource limits. If it is not used, jobs run in the system instance may inherit inappropriate limits from flux-broker.

Configuring the Network Certificate

Overlay network security requires a certificate to be distributed to all nodes. It should be readable only by the flux user. To create a new certificate, run flux-keygen(1) as the flux user, then copy the result to /etc/flux/system since the flux user will not have write access to this location:

$ sudo -u flux flux keygen /tmp/curve.cert
$ sudo mv /tmp/curve.cert /etc/flux/system/curve.cert

Do this once and then copy the certificate to the same location on the other nodes, preserving owner and mode.


The flux user only needs read access to the certificate and other files and directories under /etc/flux. Keeping these files and directories non-writable by user flux adds an extra layer of security for the system instance configuration.

Configuring the Flux System Instance

Although the security components need to be isolated, most Flux components share a common configuration space, which for the system instance is located in /etc/flux/system/conf.d/*.toml. The Flux broker for the system instance is pointed to this configuration by the systemd unit file.

Example file installed path: /etc/flux/system/conf.d/system.toml

# Flux needs to know the path to the IMP executable
imp = "/usr/libexec/flux/flux-imp"

# Allow users other than the instance owner (guests) to connect to Flux
# Optionally, root may be given "owner privileges" for convenience
allow-guest-user = true
allow-root-owner = true

# Point to shared network certificate generated flux-keygen(1).
# Define the network endpoints for Flux's tree based overlay network
# and inform Flux of the hostnames that will start flux-broker(1).
curve_cert = "/etc/flux/system/curve.cert"

default_port = 8050
default_bind = "tcp://eth0:%p"
default_connect = "tcp://%h:%p"

# Rank 0 is the TBON parent of all brokers unless explicitly set with
# parent directives.
hosts = [
   { host = "test[1-16]" },

# Speed up detection of crashed network peers (system default is around 20m)
tcp_user_timeout = "2m"

# Uncomment 'norestrict' if flux broker is constrained to system cores by
# systemd or other site policy.  This allows jobs to run on assigned cores.
# Uncomment 'exclude' to avoid scheduling jobs on certain nodes (e.g. login,
# management, or service nodes).
#norestrict = true
#exclude = "test[1-2]"

hosts = "test[1-15]"
cores = "0-7"
gpus = "0"

hosts = "test16"
cores = "0-63"
gpus = "0-1"
properties = ["fatnode"]

# Store the kvs root hash in sqlite periodically in case of broker crash.
# Recommend offline KVS garbage collection when commit threshold is reached.
checkpoint-period = "30m"
gc-threshold = 100000

# Immediately reject jobs with invalid jobspec or unsatisfiable resources
plugins = [ "jobspec", "feasibility" ]

# Remove inactive jobs from the KVS after one week.
inactive-age-limit = "7d"

# Jobs submitted without duration get a very short one
duration = "1m"

# Jobs that explicitly request more than the following limits are rejected
duration = "2h"
job-size.max.nnodes = 8
job-size.max.ncores = 32

# Configure the flux-sched (fluxion) scheduler policies
# The 'lonodex' match policy selects node-exclusive scheduling, and can be
# commented out if jobs may share nodes.
queue-policy = "easy"
match-policy = "lonodex"
match-format = "rv1_nosched"

See also: flux-config-exec(5), flux-config-access(5) flux-config-bootstrap(5), flux-config-tbon(5), flux-config-resource(5), flux-config-ingest(5), flux-config-archive(5), flux-config-job-manager(5), flux-config-policy(5), flux-config-kvs(5), flux-config-sched-fluxion-qmanager(5), flux-config-sched-fluxion-resource(5).

Configuring Resources

The Flux system instance must be configured with a static resource set. The resource.config TOML array in the example above is the preferred way to configure clusters with a resource set consisting of only nodes, cores, and GPUs.

More complex resource sets may be represented by generating a file in RFC 20 (R version 1) form with scheduler extensions using a combination of flux R encode and flux ion-R encode and then configuring resource.path to its fully-qualified file path. The details of this method are beyond the scope of this document.

When Flux is running, flux resource list shows the configured resource set and any resource properties.

Persistent Storage on Rank 0

Flux is prolific in its use of disk space to back up its key value store, proportional to the number of jobs run and the quantity of standard I/O. On your rank 0 node, ensure that the statedir directory (normally /var/lib/flux) has plenty of space and is preserved across Flux instance restarts.

The statedir directory is used for the content.sqlite file that contains content addressable storage backing the Flux key value store (KVS). The job-archive.sqlite file is also located there, if job archival is enabled.

Adding Job Prolog/Epilog Scripts

As of 0.45.0, Flux does not support a traditional job prolog/epilog which runs as root on the nodes assigned to a job before/after job execution. Flux does, however, support a job-manager prolog/epilog, which is run at the same point on rank 0 as the instance owner (typically user flux), instead of user root.

As a temporary solution, a convenience command flux perilog-run is provided which can simulate a job prolog and epilog by executing a command across the broker ranks assigned to a job from the job-manager prolog and epilog.

When using flux perilog-run to execute job prolog and epilog, the job-manager prolog/epilog feature is being used to execute a privileged prolog/epilog across the nodes/ranks assigned to a job, via the flux-security IMP “run” command support. Therefore, each of these components need to be configured, which is explained in the steps below.

  1. Configure the IMP such that it will allow the system instance user to execute a prolog and epilog script or command as root.

    allowed-users = [ "flux" ]
    path = "/etc/flux/system/prolog"
    allowed-users = [ "flux" ]
    path = "/etc/flux/system/epilog"

    By default, the IMP will set the environment variables FLUX_OWNER_USERID, FLUX_JOB_USERID, FLUX_JOB_ID, HOME and USER for the prolog and epilog processes. PATH will be set explicitly to /usr/sbin:/usr/bin:/sbin:/bin. To allow extra environment variables to be passed from the enclosing environment, use the allowed-environment key, which is an array of glob(7) patterns for acceptable environment variables, e.g.

    allowed-environment = [ "FLUX_*" ]

    will pass all FLUX_ environment variables to the IMP run commands.

  2. Configure the Flux system instance to load the job-manager perilog.so plugin, which is not active by default. This plugin enables job-manager prolog/epilog support in the instance:

    plugins = [
      { load = "perilog.so" }
  3. Configure the Flux system instance [job-manager.prolog] and [job-manager.epilog] to execute flux perilog-run with appropriate arguments to execute flux-imp run prolog and flux-imp run epilog across the ranks assigned to a job:

    command = [
       "flux", "perilog-run", "prolog",
       "-e", "/usr/libexec/flux/flux-imp,run,prolog"
    command = [
       "flux", "perilog-run", "epilog",
       "-e", "/usr/libexec/flux/flux-imp,run,epilog"

Note that the flux perilog-run command will additionally execute any scripts in /etc/flux/system/{prolog,epilog}.d on rank 0 by default as part of the job-manager prolog/epilog. Only place scripts here if there is a need to execute scripts as the instance owner (user flux) on a single rank for each job. If only traditional prolog/epilog support is required, these directories can be ignored and should be empty or nonexistent. To run scripts from a different directory, use the -d, --exec-directory option in the configured command.

See also: flux-config-job-manager(5), flux-config-security-imp(5).

Adding Job Request Validation

Jobs are submitted to Flux via a job-ingest service. This service validates all jobs before they are assigned a jobid and announced to the job manager. By default, only basic validation is done, but the validator supports plugins so that job ingest validation is configurable.

The list of available plugins can be queried via flux job-validator --list-plugins. The current list of plugins distributed with Flux is shown below:

$ flux job-validator --list-plugins
Available plugins:
feasibility           Use sched.feasibility RPC to validate job
jobspec               Python bindings based jobspec validator
require-instance      Require that all jobs are new instances of Flux
schema                Validate jobspec using jsonschema

Only the jobspec plugin is enabled by default.

In a system instance, it may be useful to also enable the feasibility and require-instance validators. This can be done by configuring the Flux system instance via the ingest TOML table, as shown in the example below:

plugins = [ "jobspec", "feasibility", "require-instance" ]

The feasibility plugin will allow the scheduler to reject jobs that are not feasible given the current resource configuration. Otherwise, these jobs are enqueued, but will have a job exception raised once the job is considered for scheduling.

The require-instance plugin rejects jobs that do not start another instance of Flux. That is, jobs are required to be submitted via tools like flux mini batch and flux mini alloc, or the equivalent. For example, with this plugin enabled, a user running flux mini run will have their job rejected with the message:

$ flux mini run -n 1000 myapp
flux-mini: ERROR: [Errno 22] Direct job submission is disabled for this instance. Please use the batch or alloc subcommands of flux-mini(1)

See also: flux-config-ingest(5).

Adding Queues

It may be useful to configure a Flux system instance with multiple queues. Each queue should be associated with a non-overlapping resource subset, identified by property name. It is good practice for queues to create a new property that has the same name as the queue.

When queues are defined, all jobs must specify a queue at submission time. If that is inconvenient, then policy.jobspec.defaults.system.queue may define a default queue.

Finally, queues can override the [policy] table on a per queue basis. This is useful for setting queue-specific limits.

Here is an example that puts these concepts together:

jobspec.defaults.system.duration = "1m"
jobspec.defaults.system.queue = "debug"

hosts = "test[1-4]"
properties = ["debug"]

hosts = "test[5-16]"
properties = ["batch"]

requires = ["debug"]
policy.limits.duration = "30m"

requires = ["batch"]
policy.limits.duration = "4h"

When named queues are configured, flux-queue(1) may be used to list them:

$ flux queue status
batch: Job submission is enabled
debug: Job submission is enabled
Scheduling is enabled

See also: flux-config-policy(5), flux-config-queues(5), flux-config-resource(5), flux-queue(1).

Policy Limits

Job duration and size are unlimited by default, or limited by the scheduler feasibility check discussed above, if configured. When policy limits are configured, the job request is compared against them after any configured jobspec defaults are set, and before the scheduler feasibility check. If the job would exceed a duration or job size policy limit, the job submission is rejected.


flux-sched 0.25.0 limitation: jobs that specify nodes but not cores may escape flux-core’s ncores policy limit, and jobs that specify cores but not nodes may escape the nnodes policy limit. The flux-sched feasibility check will eventually cover this case. Until then, be sure to set both nnodes and ncores limits when configuring job size policy limits.

Limits are global when set in the top level [policy] table. Global limits may be overridden by a policy table within a [queues] entry. Here is an example which implements duration and job size limits for two queues:

# Global defaults
jobspec.defaults.system.duration = "1m"
jobspec.defaults.system.queue = "debug"

requires = ["debug"]
policy.limits.duration = "30m"
policy.limits.job-size.max.nnodes = 2
policy.limits.job-size.max.ncores = 16

requires = ["batch"]
policy.limits.duration = "8h"
policy.limits.job-size.max.nnodes = 16
policy.limits.job-size.max.ncores = 128

See also: flux-config-policy(5).

Use PAM to Restrict Access to Compute Nodes

If Pluggable Authentication Modules (PAM) are in use within a cluster, it may be convenient to use the pam_flux.so account module to configure a PAM stack that denies users access to compute nodes unless they have a job running there.

Install the flux-pam package to make the pam_flux.so module available to be added to one or more PAM stacks, e.g.

account  sufficient pam_flux.so

Day to day administration

Starting Flux

Systemd may be configured to start Flux automatically at boot time, as long as the network that carries its overlay network will be available at that time. Alternatively, Flux may be started manually, e.g.

$ sudo pdsh -w fluke[3,108,6-103] sudo systemctl start flux

Flux brokers may be started in any order, but they won’t come online until their parent in the tree based overlay network is available.

If Flux was not shut down properly, for example if the rank 0 broker crashed or was killed, then Flux starts in a safe mode with job submission and scheduling disabled. flux-uptime(1) shows the general state of Flux, and flux-startlog(1) prints a record of Flux starts and stops, including any crashes.

Stopping Flux

The full Flux system instance may be temporarily stopped by running the following on the rank 0 node:

$ sudo flux shutdown

This kills any running jobs, but preserves job history and the queue of jobs that have been submitted but have not yet allocated resources. This state is held in the content.sqlite that was configured above. See also flux-shutdown(1).


flux-shutdown --gc should be used from time to time to perform offline KVS garbage collection. This, in conjunction with configuring inactive job purging, keeps the size of the content.sqlite database in check and improves Flux startup time.

The brokers on other nodes will automatically shut down in response, then respawn, awaiting the return of the rank 0 broker.

To shut down a single node running Flux, simply run

$ sudo systemctl stop flux

on that node.

Configuration update

After changing flux broker or module specific configuration in the TOML files under /etc/flux, the configuration may be reloaded with

$ sudo systemctl reload flux

on each rank where the configuration needs to be updated. The broker will reread all configuration files and notify modules that configuration has been updated.

Configuration which applies to the flux-imp or job shell will be reread at the time of the next job execution, since these components are executed at job launch.


Many configuration changes have no effect until the Flux broker restarts. This should be assumed unless otherwise noted. See flux-config(5) for more information.

Viewing resource status

Flux offers two different utilities to query the current resource state.

flux resource status is an administrative command which lists ranks which are available, online, offline, excluded, or drained along with their corresponding node names. By default, sets which have 0 members are not displayed, e.g.

$ flux resource status
    avail     15 1-15            fluke[26-40]
    drain      1 0               fluke25

To list a set of states explicitly, use the --states option: (Run --states=help to get a list of valid states)

$ flux resource status --states=offline,exclude
  offline      0
  exclude      0

This option is useful to get a list of ranks or hostnames in a given state. For example, the following command fetches the hostlist for all resources configured in a Flux instance:

$ flux resource status -s all -no {nodelist}

In contrast to flux resource status, the flux resource list command lists the scheduler’s view of available resources. This command shows the free, allocated, and unavailable (down) resources, and includes nodes, cores, and gpus at this time:

$ flux resource list
     free     15       60        0 fluke[26-40]
allocated      0        0        0
     down      1        4        0 fluke25

With -v, flux resource list will show a finer grained list of resources in each state, instead of a nodelist:

$ flux resource list -v
      free     15       60        0 rank[1-15]/core[0-3]
 allocated      0        0        0
      down      1        4        0 rank0/core[0-3]

Draining resources

Resources may be temporarily removed from scheduling via the flux resource drain command. Currently, resources may only be drained at the granularity of a node, represented by its hostname or broker rank, for example:

$ sudo flux resource drain fluke7 node is fubar
$ sudo flux resource drain
TIMESTAMP            STATE    RANK     REASON                         NODELIST
2020-12-16T09:00:25  draining 2        node is fubar                  fluke7

Any work running on the “draining” node is allowed to complete normally. Once there is nothing running on the node its state changes to “drained”:

$ sudo flux resource drain
TIMESTAMP            STATE    RANK     REASON                         NODELIST
2020-12-16T09:00:25  drained  2        node is fubar                  fluke7

To return drained resources use flux resource undrain:

$ sudo flux resource undrain fluke7
$ sudo flux resource drain
TIMESTAMP            STATE    RANK     REASON                         NODELIST

Managing the Flux queue

The queue of jobs is managed by the flux job-manager, which in turn makes allocation requests for jobs in priority order to the scheduler. This queue can be managed using the flux-queue command.

Usage: flux-queue [OPTIONS] COMMAND ARGS
  -h, --help             Display this message.

Common commands from flux-queue:
   enable          Enable job submission
   disable         Disable job submission
   start           Start scheduling
   stop            Stop scheduling
   status          Get queue status
   drain           Wait for queue to become empty.
   idle            Wait for queue to become idle.

The queue may be listed with the flux-jobs(1) command.

Disabling job submission

By default, the queue is enabled, meaning that jobs can be submitted into the system. To disable job submission, e..g to prepare the system for a shutdown, use flux queue disable. To restore queue access use flux queue enable.

Stopping resource allocation

The queue may also be stopped with flux queue stop, which disables further allocation requests from the job-manager to the scheduler. This allows jobs to be submitted, but stops new jobs from being scheduled. To restore scheduling use flux queue start.

Flux queue idle and drain

The flux queue drain and flux queue idle commands can be used to wait for the queue to enter a given state. This may be useful when preparing the system for a downtime.

The queue is considered drained when there are no more active jobs. That is, all jobs have completed and there are no pending jobs. flux queue drain is most useful when the queue is disabled .

The queue is “idle” when there are no jobs in the RUN or CLEANUP state. In the idle state, jobs may still be pending. flux queue idle is most useful when the queue is stopped.

To query the current status of the queue use the flux queue status command:

$ flux queue status -v
flux-queue: Job submission is enabled
flux-queue: Scheduling is enabled
flux-queue: 2 alloc requests queued
flux-queue: 1 alloc requests pending to scheduler
flux-queue: 0 free requests pending to scheduler
flux-queue: 4 running jobs

Managing Flux jobs

Expediting/Holding jobs

To expedite or hold a job, set its urgency to the special values EXPEDITE or HOLD.

$ flux job urgency ƒAiVi2Sj EXPEDITE
$ flux job urgency ƒAiVi2Sj HOLD

Canceling jobs

An active job may be canceled via the flux job cancel command. An instance owner may cancel any job, while a guest may only cancel their own jobs.

All active jobs may be canceled with flux job cancelall. By default this command will only print the number of jobs that would be canceled. To force cancellation of all matched jobs, the -f, --force option must be used:

$ flux job cancelall
flux-job: Command matched 5 jobs (-f to confirm)
$ flux job cancelall -f
flux-job: Canceled 5 jobs (0 errors)

The set of jobs matched by the cancelall command may also be restricted via the -s, --states=STATES and -u, --user=USER options.

Software update

Flux will eventually support rolling software upgrades, but prior to major release 1, Flux software release versions should not be assumed to inter-operate. Furthermore, at this early stage, Flux software components (e.g. flux-core, flux-sched, flux-security, and flux-accounting) should only only be installed in recommended combinations.


Mismatched broker versions are detected as brokers attempt to join the instance. The version is currently required to match exactly.


Overlay network

The tree-based overlay network interconnects brokers of the system instance. The current status of the overlay subtree at any rank can be shown with:

$ flux overlay status -r RANK

The possible status values are:


Node is online and no children are in partial, offline, degraded, or lost state.


Node is online, and some children are in partial or offline state; no children are in degraded or lost state.


Node is online, and some children are in degraded or lost state.


Node has gone missing, from the parent perspective.


Node has not yet joined the instance, or has been cleanly shut down.

Note that the RANK argument is where the request will be sent, not necessarily the rank whose status is of interest. Parents track the status of their children, so a good approach when something is wrong to start with rank 0 (the default). The following options can be used to ask rank 0 for a detailed listing:

$ flux overlay status
0 fluke62: degraded
├─ 1 fluke63: full
│  ├─ 3 fluke65: full
│  │  ├─ 7 fluke70: full
│  │  └─ 8 fluke71: full
│  └─ 4 fluke67: full
│     ├─ 9 fluke72: full
│     └─ 10 fluke73: full
└─ 2 fluke64: degraded
   ├─ 5 fluke68: full
   │  ├─ 11 fluke74: full
   │  └─ 12 fluke75: full
   └─ 6 fluke69: degraded
      ├─ 13 fluke76: full
      └─ 14 fluke77: lost

To determine if a broker is reachable from the current rank, use:

$ flux ping RANK

A broker that is not responding but is not shown as lost or offline by flux overlay status may be forcibly detached from the overlay network with:

$ flux overlay disconnect RANK

However, before doing that, it may be useful to see if a broker acting as a router to that node is actually the problem. The overlay parent of RANK may be listed with

$ flux overlay parentof RANK

Using flux ping and flux overlay parentof iteratively, one should be able to isolate the problem rank.

See also flux-overlay(1), flux-ping(1).

Systemd journal

Flux brokers log information to standard error, which is normally captured by the systemd journal. It may be useful to look at this log when diagnosing a problem on a particular node:

$ journalctl -u flux
Sep 14 09:53:12 sun1 systemd[1]: Starting Flux message broker...
Sep 14 09:53:12 sun1 systemd[1]: Started Flux message broker.
Sep 14 09:53:12 sun1 flux[23182]: broker.info[2]: start: none->join 0.0162958s
Sep 14 09:53:54 sun1 flux[23182]: broker.info[2]: parent-ready: join->init 41.8603s
Sep 14 09:53:54 sun1 flux[23182]: broker.info[2]: rc1.0: running /etc/flux/rc1.d/01-enclosing-instance
Sep 14 09:53:54 sun1 flux[23182]: broker.info[2]: rc1.0: /bin/sh -c /etc/flux/rc1 Exited (rc=0) 0.4s
Sep 14 09:53:54 sun1 flux[23182]: broker.info[2]: rc1-success: init->quorum 0.414207s
Sep 14 09:53:54 sun1 flux[23182]: broker.info[2]: quorum-full: quorum->run 9.3847e-05s

Broker log buffer

The rank 0 broker accumulates log information for the full instance in a circular buffer. For some problems, it may be useful to view this log:

$ sudo flux dmesg |tail
2020-09-14T19:38:38.047025Z sched-simple.debug[0]: free: rank1/core0
2020-09-14T19:38:41.600670Z sched-simple.debug[0]: req: 6115337007267840: spec={0,1,1} duration=0.0
2020-09-14T19:38:41.600791Z sched-simple.debug[0]: alloc: 6115337007267840: rank1/core0
2020-09-14T19:38:41.703252Z sched-simple.debug[0]: free: rank1/core0
2020-09-14T19:38:46.588157Z job-ingest.debug[0]: validate-jobspec.py: inactivity timeout