clush is a program for executing commands in parallel on a cluster and for
gathering their results. It can execute commands interactively or can be used
within shell scripts and other applications. It is a partial front-end to the
Task class of the ClusterShell library (cf. Structure of Task).
clush currently makes use of the Ssh worker of ClusterShell that only
requires ssh(1) (we tested with OpenSSH SSH client).
Some features of clush command line tool are:
- two modes of parallel cluster commands execution:
- flat mode: sliding window of local or remote (eg. ssh(1)) commands
- tree mode: commands propagated to the targets through a tree of pre-configured gateways; gateways are then using a sliding window of local or ssh(1) commands to reach the targets (if the target count per gateway is greater than the fanout value)
- smart display of command results (integrated output gathering, sorting by node, nodeset or node groups)
- standard input redirection to remote nodes
- files copying in parallel
- pdsh  options backward compatibility
clush can be started non-interactively to run a shell command, or can be invoked as an interactive shell. Both modes are discussed here (clush-oneshot clush-interactive).
Target and filter nodes¶
clush offers different ways to select or filter target nodes through command line options or files containing a list of hosts.
Command line options¶
-w option allows you to specify remote hosts by using ClusterShell
NodeSet syntax, including the node groups @group special syntax
(cf. Node group expression rules) and the Extended String Patterns syntax (see
Extended String Pattern) to benefits from
basic arithmetics (like
@Agroup&@Bgroup). Additionally, the
allows you to exclude nodes from remote hosts list (the same NodeSet syntax
can be used here). Nodes exclusion has priority over nodes addition.
Using node groups¶
If you have ClusterShell node groups configured on your
cluster, any node group syntax may be used in place of nodes for
$ clush -w @rhel6 cat /proc/loadavg node26: 0.02 0.01 0.00 1/202 23042
For pdsh backward compatibility, clush supports two
options to respectively select and exclude nodes group(s), but only specified
by omitting any “@” group prefix (see example below). In general, though, it
is advised to use the @-prefixed group syntax as the non-prefixed notation
is only recognized by clush but not by other tools like nodeset.
$ clush -g rhel6 cat /proc/loadavg node26: 0.02 0.01 0.00 1/202 23033
Selecting all nodes¶
The special option
-a (without argument) can be used to select all
nodes, in the sense of ClusterShell node groups (see
node groups configuration for more details on special
all external shell command upcall). If not properly configured, the
-a option may lead to a runtime error like:
clush: External error: Not enough working external calls (all, or map + list) defined to get all node
Picking node(s) at random¶
--pick with a maximum number of nodes you wish to pick randomly from
the targeted node set. clush will then run only on selected node(s). The
following example will run a script on a single random node picked from the
$ clush -w @compute --pick=1 ./nonreg-single-client-fs-io.sh
--machinefile) may be used to specify a
path to a file containing a list of single hosts, node sets or node groups,
separated by spaces and lines. It may be specified multiple times (one per
$ clush --hostfile ./host_file -b systemctl is-enabled httpd
This option has been added as backward compatibility with other parallel shell
tools. Indeed, ClusterShell provides a preferred way to provision node sets
from node group sources and flat files to all cluster tools using
NodeSet (including clush). Please see node groups
-d to see resulting node sets from host
Flat execution mode¶
The default execution mode is to launch commands (local or remote) in parallel, up to a certain limit fixed by the fanout value, which is the number of child processes allowed to run at a time. This “sliding window” of active commands is a common technique used on large clusters to conserve resources on the initiating host, while allowing some commands to time out. If used with ssh(1), this does actually limit the number of concurrent ssh connections.
Fanout (sliding window)¶
Indeed, it is sometimes useful to change the fanout value for a specific command, for example to avoid flooding a remote service with concurrent requests generated by that actual command.
The following example will launch up to ten puppet commands at a time on the node group named @compute:
$ clush -w @compute -f 10 puppet agent -t
If the fanout value is set to 1, commands are executed sequentially:
$ clush -w node[40-42] -f 1 'date +%s; sleep 1' node40: 1505366138 node41: 1505366139 node42: 1505366140
Tree execution mode¶
ClusterShell’s tree execution mode is a major horizontal scalability improvement by providing a hierarchical command propagation scheme.
The Tree mode of ClusterShell has been the subject of this paper presented at the Ottawa Linux Symposium Conference in 2012 and at the PyHPC 2013 workshop in Denver, USA.
The diagram below illustrates the hierarchical command propagation principle with a head node, gateways (GW) and target nodes:
.-----------. | Head node | '-----------' /|\ .------------' | '--.-----------. / | \ \ .-----. .-----. \ .-----. | GW1 | | GW2 | \ | GW3 | '-----' '-----' \ '-----' /|\ /|\ \ |\ .-' | '-. .-' | '-. \ | '---. / | \ / | \ \ | \ .---. .---. .---. .---. .---. .---. .---. .---. .-----. '---' '---' '---' '---' '---' '---' '---' '---' | GW4 | target nodes '-----' | ...
The Tree mode is implemented at the library level, so that all applications using ClusterShell may benefits from it. However, this section describes how to use the tree mode with the clush command only.
The system-wide library configuration file /etc/clustershell/topology.conf defines the routes of default command propagation tree. It is recommended that all connections between parent and children nodes are carefully pre-configured, for example, to avoid any SSH warnings when connecting (if using the default SSH remote worker, of course).
The content of the topology.conf file should look like this:
[routes] rio0: rio[10-13] rio[10-11]: rio[100-240] rio[12-13]: rio[300-440]
This file defines the following topology graph:
rio0 |- rio[10-11] | `- rio[100-240] `- rio[12-13] `- rio[300-440]
At runtime, ClusterShell will pick an initial propagation tree from this topology graph definition and the current root node. Multiple admin/root nodes may be defined in the file.
The algorithm used in Tree mode does not rely on gateway system hostnames anymore. In topology.conf, just use the hosts or aliases needed to connect to each node.
Enabling tree mode¶
Since version 1.7, the tree mode is enabled by default when a configuration
file is present. When the configuration file
/etc/clustershell/topology.conf exists, clush will use it by default for
target nodes that are defined there. The topology file path can be changed
--topology command line option.
clush -d (debug option), clush will display an ASCII
representation of the initial propagation tree used. This is useful when
working on Tree mode configuration.
Enabling tree mode should be as much transparent as possible to the end user.
Most clush options, including options defined in
clush.conf or specified using
options) are propagated to the gateways and taken into account there.
Tree mode specific options¶
--remote=yes|no command line option controls the remote execution
Default is yes, that will make clush establish connections up to the leaf nodes using a distant worker like ssh.
Changing it to no will make clush establish connections up to the leaf parent nodes only, then the commands are executed locally on the gateways (like if it would be with
--worker=execon the gateways themselves). This execution mode allows users to schedule remote commands on gateways that take a node as an argument. On large clusters, this is useful to spread the load and resources used of one-shot monitoring, IPMI, or other commands on gateways. A simple example of use is:
$ clush -w node[100-199] --remote=no /usr/sbin/ipmipower -h %h-ipmi -s
This command is also valid if you don’t have any tree configured, because in that case,
--remote=nois an alias of
--grooming command line option allows users to change the grooming
delay (float, in seconds). This feature allows gateways to aggregate responses
received within a certain timeframe before transmitting them back to the root
node in a batch fashion. This contributes to reducing the load on the root
node by delegating the first steps of this CPU intensive task to the gateways.
ClusterShell uses a “sliding window” or fanout of processes to avoid too many concurrent connections and to conserve resources on the initiating hosts. See Flat execution mode for more details about this.
In tree mode, the same fanout value is used on the head node and on each gateway. That is, if the fanout is 16, each gateway will initiate up to 16 connections to their target nodes at the same time.
This is likely to change in the future, as it makes the fanout behaviour different if you are using the tree mode or not. For example, some administrators are using a fanout value of 1 to “sequentialize” a command on the cluster. In tree mode, please note that in that case, each gateway will be able to run a command at the same time.
Remote Python executable¶
You must use the same major version of Python on the gateways and the root node. By default, the same python executable name than the one used on the root node will be used to launch the gateways, that is, python or python3 (using relative path for added flexibility). You may override the selection of the remote Python interpreter by defining the following environment variable:
$ export CLUSTERSHELL_GW_PYTHON_EXECUTABLE=/path/to/python3
It is highly recommended to have the same Python interpeter installed on all gateways and the root node.
Debugging Tree mode¶
To debug Tree mode, you can define the following environment variable before running clush (or any other applications using ClusterShell):
$ export CLUSTERSHELL_GW_LOG_LEVEL=DEBUG (default value is INFO) $ export CLUSTERSHELL_GW_LOG_DIR=/tmp (default value is /tmp)
This will generate log files of the form
Non-interactive (or one-shot) mode¶
When clush is started non-interactively, the command is executed on the specified remote hosts in parallel (given the current fanout value and the number of commands to execute (see fanout library settings in Configuring the Task object).
Output gathering options¶
--dshbak is specified, clush waits for command
completion while displaying a progress indicator and
then displays gathered output results. If standard output is redirected to a
file, clush detects it and disable any progress indicator.
clush will only consolidate identical command outputs if the command return codes are also the same.
The following is a simple example of clush command used to execute
-r on node40, node41 and node42, wait for their completion and finally
display digested output results:
$ clush -b -w node[40-42] uname -r --------------- node[40-42] --------------- 220.127.116.11-45.fc14.x86_64
It is common to cancel such command execution because a node is hang. When using pdsh and dshbak, due to the pipe, all nodes output will be lost, even if all nodes have successfully run the command. When you hit CTRL-C with clush, the task is canceled but received output is not lost:
$ clush -b -w node[1-5] uname -r Warning: Caught keyboard interrupt! --------------- node[2-4] (3) --------------- 18.104.22.168-145.fc11 --------------- node5 --------------- 2.6.18-164.11.1.el5 Keyboard interrupt (node1 did not complete).
Performing diff of cluster-wide outputs¶
Since version 1.6, you can use the
--diff clush option to show
differences between common outputs. This feature is implemented using Python
unified diff. This special option implies
-b (gather common stdout
outputs) but you don’t need to specify it. Example:
$ clush -w node[40-42] --diff dmidecode -s bios-version --- node[40,42] (2) +++ node41 @@ -1,1 +1,1 @@ -1.0.5S56 +1.1c
A nodeset is automatically selected as the “reference nodeset” according to these criteria:
- lowest command return code (to discard failed commands)
- largest nodeset with the same output result
- otherwise the first nodeset is taken (ordered (1) by name and (2) lowest range indexes)
Standard input bindings¶
Unless the option
-n) is specified, clush detects when
its standard input is connected to a terminal (as determined by isatty(3)).
If actually connected to a terminal, clush listens to standard input when
commands are running, waiting for an Enter key press. Doing so will display the
status of current nodes. If standard input is not connected to a terminal, and
unless the option
-n) is specified, clush binds the
standard input of the remote commands to its own standard input, allowing
scripting methods like:
$ echo foo | clush -w node[40-42] -b cat --------------- node[40-42] --------------- foo
Another stdin-bound clush usage example:
$ ssh node10 'ls /etc/yum.repos.d/*.repo' | clush -w node[11-14] -b xargs ls --------------- node[11-14] (4) --------------- /etc/yum.repos.d/cobbler-config.repo
-n) in the same way you would use
to disable standard input. Indeed, if this option is set, EOF is sent at
first read, as if stdin were actually connected to /dev/null.
In output gathering mode, clush will display a live
progress indicator as a simple but convenient way to follow the completion of
parallel commands. It can be disabled just by using the
options. The progress indicator will appear after 1 to 2 seconds and should
look like this:
If writing is performed to clush standard input, like in
clush, the live progress indicator will display the global bandwidth of data
written to the target nodes.
Finally, the special option
--progress can be used to force the display of
the live progress indicator. Using this option may interfere with some command
outputs, but it can be useful when using stdin while remote commands are
silent. As an example, the following command will copy a local file to
node[1-3] and display the global write bandwidth to the target nodes:
$ dd if=/path/to/local/file | clush -w node[1-3] --progress 'dd of=/path/to/remote/file' clush: 0/3 write: 212.27 MiB/s
If a command is not specified, clush runs interactively. In this mode, clush uses the GNU readline library to read command lines from the terminal. Readline provides commands for searching through the command history for lines containing a specified string. For instance, you can type Control-R to search in the history for the next entry matching the search string typed so far.
Single-character interactive commands¶
clush also recognizes special single-character prefixes that allows the user to see and modify the current nodeset (the nodes where the commands are executed). These single-character interactive commands are detailed below:
|Interactive special commands||Comment|
||show current nodeset|
||add nodes to current nodeset|
||remove nodes from current nodeset|
||set current nodeset|
||toggle the ouput format (gathered or standard mode)|
To leave an interactive session, type
quit or Control-D. As of version
1.6, it is not possible to cancel a command while staying in clush
interactive session: for instance, Control-C is not supported and will abort
current clush interactive command (see ticket #166).
Example of clush interactive session:
$ clush -w node[11-14] -b Enter 'quit' to leave this interactive mode Working with nodes: node[11-14] clush> uname --------------- node[11-14] (4) --------------- Linux clush> !pwd LOCAL: /root clush> -node[11,13] Working with nodes: node[12,14] clush> uname --------------- node[12,14] (2) --------------- Linux clush>
The interactive mode and commands described above are subject to change and improvements in future releases. Feel free to open an enhancement ticket if you use the interactive mode and have some suggestions.
File copying mode¶
When clush is started with the
--copy option, it will
attempt to copy specified file and/or directory to the provided target cluster
nodes. If the
--dest option is specified, it will put the copied files
or directory there.
Here are some examples of file copying with clush:
$ clush -v -w node[11-12] --copy /tmp/foo `/tmp/foo' -> node[11-12]:`/tmp' $ clush -v -w node[11-12] --copy /tmp/foo /tmp/bar `/tmp/bar' -> aury[11-12]:`/tmp' `/tmp/foo' -> aury[11-12]:`/tmp' $ clush -v -w node[11-12] --copy /tmp/foo --dest /var/tmp/ `/tmp/foo' -> node[11-12]:`/var/tmp/'
Reverse file copying mode¶
When clush is started with the
--rcopy option, it will attempt to
retrieve specified file and/or directory from provided cluster nodes. If the
--dest option is specified, it must be a directory path where the files
will be stored with their hostname appended. If the destination path is not
specified, it will take the first file or dir basename directory as the local
$ clush -v -w node[11-12] --rcopy /tmp/foo node[11-12]:`/tmp/foo' -> `/tmp' $ ls /tmp/foo.* /tmp/foo.node11 /tmp/foo.node12
Overriding clush.conf settings¶
clush default settings are found in a configuration described in
clush configuration. To override any settings, use the
--option command line option (or
-O for the shorter version), and
repeat as needed. Here is a simple example to disable the use colors in the
output nodeset header:
$ clush -O color=never -w node[11-12] -b echo ok --------------- node[11-12] (2) --------------- ok
By default, clush is using the default library worker configuration when running commands or copying files. In most cases, this is ssh (See Changing default worker for default worker selection).
Worker selection can be performed at runtime thanks to
line option (or
-R for the shorter version in order to be compatible with
pdsh remote command selection option):
$ clush -w node[11-12] --worker=rsh echo ok node11: ok node12: ok
By default, ClusterShell supports the following worker identifiers:
exec: this local worker supports parallel command execution, doesn’t rely on any external tool and provides command line placeholders described below:
%hostare substitued with each target hostname
%hostsis substitued with the full target nodeset
%rankare substitued with the remote rank (0 to n-1)
For example, the following would request the exec worker to locally run multiple ipmitool commands across the hosts foo[0-10] and automatically aggregate output results (-b):
$ clush -R exec -w foo[0-10] -b ipmitool -H %h-ipmi chassis power status --------------- foo[0-10] (11) --------------- Chassis Power is on
rsh: remote worker based on rsh
ssh: remote worker based on ssh (default)
pdsh: remote worker based on pdsh that requires pdsh to be installed; doesn’t provide write support (eg. you cannot
cat file | clush --worker pdsh); it is primarily an 1-to-n worker example.
|||LLNL parallel remote shell utility (https://computing.llnl.gov/linux/pdsh.html)|