NAME¶

remote_startup - the Grid Engine remote startup mechanism

DESCRIPTION¶

Grid Engine supports several commands to facilitate interactive commands or remote startup of a tightly integrated parallel job. Each command can be set up with qconf(1) (option -sconf) to use different daemons and commands to start the final session. Different startup methods can therefore contain different daemons and commands, and are not related to other startup methods in any way, although it is often desirable to have the same communication method for all startup methods.

Each method requires a separate instance of the communication daemon to be started by sge_shepherd(8) for each job, which thus must use a randomly-chosen port, to which the client will try to connect. This is necessary to support tight integration, ensuring that everything for a given job stays in the same process tree and can be properly controlled and accounted by Grid Engine.

QLOGIN¶

An interactive qlogin session invoked by qlogin(1) will show up in qstat with the default name QLOGIN unless changed by the -N name option. The two entries qlogin_daemon and qlogin_command are responsible for establishing the communication to start such a session.

The default is the value builtin, which will trigger an SGE internal communication method:

In cases where you want a different communication method, it can also be set up for the formerly-used communication method based on telnet. Despite the fact that telnet is used, there is no need to have telnetd running all the time - SGE will start a unique one for each job, even when it's from the same user, and telnet can stay disabled as a system service (whether under init or inetd). Only the file /etc/hosts.equiv needs to contain the name of the machines from where a qlogin(1) should be allowed. This is often the head node of a cluster, or particular submit machines. Using telnet, the traffic is not encrypted, but that may be reasonable on a private subnet for the cluster, especially if SGE's CSP security isn't used to secure the system generally.

To achieve this:

must be defined in sge_conf(5). The defined qlogin_command will then be called with two additional parameters: ‘HOST’, and ‘PORT’ in exactly that order, which refer to the machine the qlogin_command should address and the port to be used.

This can also be used to set up secure communication using SSH (which can also provide X and credential forwarding, as well as compression). In this case a small wrapper must be implemented, whose sole purpose is to swap the two given arguments and prepend -p to the port argument. A suitable one is installed as $SGE_ROOT/util/resources/wrappers/qlogin_wrapper:

QRLOGIN¶

An interactive qrlogin session invoked by qrsh(1) without a command will show up in qstat having the default name QRLOGIN unless changed by the -N name option. The two entries rlogin_daemon and rlogin_command are responsible for establishing the command to start such a session.

The default is the value builtin, which will trigger an SGE internal communication method:

In cases where you want a different communication method, it can follow the formerly-used communication method based on rlogin. As for telnet, rlogin can stay disabled as a system service (whether under init or inetd). Only the file /etc/hosts.equiv needs to contain the name of the machines from where a qrsh(1) should be allowed. This is often the head node of a cluster, or particular submit machines. Using rlogin, the traffic is not encrypted, but that may be reasonable on a private subnet for the cluster, especially if SGE's CSP security isn't used to secure the system generally.

To achieve this:

must be defined in sge_conf(5). The value of $SGE_ROOT must be replaced by the root of the installation, and $ARC must be replaced by the particular platform architecture, as use of environment variables is not implemented for these entries. When the cluster is homogeneous, it can be set to e.g. ‘lx-amd64’ or ‘lx-x86’. In a heterogeneous cluster local configurations need to be defined, where preferably the minority of machines will get local configurations.

The defined rlogin_command will then be called with three additional parameters: ‘-p’,‘PORT’, and ‘HOST’ in exactly that order, which refer to the machine the rlogin_command should address, and the port to be used.

This can also be used to set up a secure communication using SSH:

QRSH¶

An interactive session for a remote command invoked by qrsh(1) with a command will show up in qstat by default with name of the command issued, unless changed by the -N name option. The two entries rsh_daemon and rsh_command are responsible for establishing the communication to start such a session. This startup method will also be used by the master task of a tightly integrated parallel job to start slave processes on other granted exechosts.

The default is the value builtin, which will trigger an SGE internal communication method:

In cases where you want a different communication method, it can also be set up for the formerly-used communication method based on rsh. As for telnet, rsh can stay disabled as a system service (whether under init or inetd). Only the file /etc/hosts.equiv needs to contain the name of the machines from where a qrsh(1) should be allowed. This is often the head node of a cluster, or particular submit machines. Using rsh, the traffic is not encrypted, but that may be reasonable on a private subnet for the cluster, especially if SGE's CSP security isn't used to secure the system generally.

To achieve this:

must be defined in sge_conf(5). The value of $SGE_ROOT must be replaced by the root of the installation, and $ARC must be replaced by the particular platform architecture, as use of environment variables is not implemented for these entries. When the cluster is homogeneous, it can be set to e.g. ‘lx-amd64’ or ‘lx-x86’. In a heterogeneous cluster local configurations need to be defined, where preferably the minority of machines will get local configurations.

The defined rsh_command will then be called with four additional parameters: ‘-n’, ‘-p’, ‘PORT’, and ‘HOST’ in exactly that order, which refer to the machine the rsh_command should address and the port to be used.

This can also be used to set up a secure communication using SSH:

LOCAL CONFIGURATIONS OF EXECHOSTS¶

It is important to note that the communication method set up for one particular startup method must match at each end. This can either be achieved by using only a global configuration, or carefully setting up local configurations for the exechosts involved. Whether or not local configurations exist, which must be taken care of, can be checked with qconf -sconfl.

As a general rule, for setting up a communication method between a machine A (where the command is issued) and a machine B (where the daemon is started) it must be guaranteed that the:

matches

This way it is also possible to use different communication methods, depending whether a connection from A to B is invoked, or from B to A.

RESTRICTIONS¶

For all three communication methods, a direct connection between the target and the source machine where the particular command was issued must exist. This can also be implemented using TCP/IP forwarding, but will usually fail if one machine is behind NAT which will mangle the machines' addresses. The communication methods won't work with simple firewalling of the exec hosts since the methods use a random port. It may be possible to set up application-specific firewalling, if necessary, or to wrap the methods and start an SSH tunnel on the port specified for each communication instance.

The builtin method does not support forwarding of X graphics from the compute nodes, or GSSAPI tokens to them. If you need that for any of the remote methods, you will want to set up SSH communication instead.

SSH AUTHENTICATION¶

To allow the SSH setup explained above to work, the user must be authenticated without the use of a passphrase. While entering a passphrase would still work for interactive commands, it will fail in case of a tightly integrated parallel job, where the master process tries to start a slave process on another exechost.

You can set up passphraseless SSH keys, although this is discouraged. A simpler and global working setup is to use host-based authentication ⟨URL: http://arc.liv.ac.uk/SGE/howto/hostbased-ssh.html ⟩ for the machines inside the cluster.

SSH TIGHT INTEGRATION¶

To have a tight integration of SSH into SGE, the started sshd needs an additional group ID to be attached. With this additional group ID, SGE is able to record the resource consumption and computing time in a correct way. Also a qdel of such a job will be able to succeed.

Such a tight SSH integration can be achieved by two means:

Use of PAM

The easiest way on supported platforms (at least GNU/Linux): a pam(7) module pam_sge-qrsh-setup(8) is available for use with the system ssh; it attaches the necessary additional group ID to the started process to provide tight integration. See also the workshop paper ⟨URL: http://gridengine.org/assets/static/ws2007/K5SGE.pdf ⟩.

Recompile Grid Engine with ./aimk -tight-ssh ...

The source of Grid Engine contains the necessary additions to compile a modified sshd, which will honor the additional group ID and attach it also to the started process. It's necessary to provide the source of OpenSSH in the directory 3rd_party inside $SGE_ROOT having a plain name ‘openssh’. Inside this directory the file sshd.c needs to be patched:

RESTRICTING ACCESS¶

With the builtin method in use, there is no need to allow direct access for normal users to compute nodes with ssh etc. However, you may want to allow users to access the nodes for debugging. If you don't want to over-subscribe the nodes, so that qrsh etc. can be used for access, you can use PAM to restrict access for a user only to the nodes on which they have a running job, so as to minimize interference with other others.

There are two possible ways. The cleanest uses pam_sge_authorize(8). Otherwise you can use generic PAM modules, such as pam_limits(8) or pam_access(8), with modifications to their configuration set up and taken down in the job prolog and epilog respectively. See, for instance, a user list message ⟨URL: http://gridengine.markmail.org/message/mu3i7haeshlevu6q?q=282211 ⟩, and other examples of similar prolog/epilog scripts provided with locking in the pam_authuser contribution in the Torque distribution.

SECURITY¶

See the notes above concerning security of the communication channel.

EXAMPLES¶

Using SSH with the PAM module, forcing tty allocation, and preventing the delegation of GSSAPI credentials to the compute nodes:

Old-style method, using telnet and rlogin:

FILES¶

$SGE_ROOT/util/resources/wrappers/qlogin_wrapper

SSH-based wrapper for qlogin (see above)

SEE ALSO¶

qconf(1), qlogin(1), qrsh(1), sge_conf(5), pam_sge_authorize(8), pam_sge-qrsh-setup(8), Grid Engine-specific remote programs ⟨URL: http://arc.liv.ac.uk/repos/darcs/sge/source/3rdparty/remote/remote.html ⟩.

AUTHOR¶

Man page written by Reuti, partly based on Sun material. Some additions by Dave Love.

COPYRIGHT¶

See sge_intro(1) for a full statement of rights and permissions.