SCORED(8)
MAINTENANCE COMMANDS
SCORED(8)
NAME
scored - global operating system

SYNOPSIS
scored [-ts time_slice] [-syslog address[:port]] [-sysmon address[:port]] [-server server_node] [-network network_name] [-operator username] [-session session-no] [-v] [-restart] [-reset|-resetall]

DESCRIPTION
SCore-D is a user-level parallel operating system providing a single system image for a cluster. scored is a scout command, this means scored command must be invoked in a scout(1) environment, and never terminates by itself.

scored will run within the SCOUT environment assuming the host group of the environment is a cluster. No need of scout command prefix to invoke scored, since scored can only run within a scout environment. scored is also invoked implicitly when users run their parallel programs with the scrun(1) command. In this case, scored runs in single-user mode. scored runs in multi-user mode when invoked explicitly.

scored manages various cluster resources, such as processors and networks. Scheduling user parallel jobs is one of the major task of scored. scored schedules user parallel jobs in a Time Space Sharing Scheduling fashion; multiple user parallel jobs can be multiplexed in time and space domains to obtain higher throughput, and shorter response time. Because of this scheduling strategy to balance the cluster-wide load, users cannot specify where to run their parallel jobs.

When SCore-D is running in multi-user mode, and if the file named scored.rc is readable from the SCore-D server host, then the file is read for initial setting. The file is assumed to contain a series of SCore-D console commands. The scored.rc file is searched firstly in the install directory (/opt/score/deploy/ in default configuration).

scored fork()'s and exec()'s user processes on compute hosts. scored (multi-user mode) must be executed as the root, so that it can set the user ID and group IDs of user processes. User's program executable files are copied by scored and scrun to the cluster hosts. Thus executable files do not need to be located in a network file system. However, user executable files must be readable for copying.

Users executable files are copied into the SCore-D directory. The SCore-D directory is /var/scored/multiuser/ in multi-user mode, or /var/tmp/scored/singleuser/<user-ID>/ in sigle-user mode. Here, <user-ID> is the user ID number. One of these directories may be created on the cluster hosts when scored is firstly invoked. The user files are removed when a parallel job is terminated. However, when scored running in sigle-user mode carshes for some reason, there can be the case in which /var/tmp/scored/singleuser/<user-ID> is not removed. When this happens, scored may fail to startup with the message "SCore-D directory (/var/tmp/scored/singleuser/1234) already exists." To avoid this, delete the directory before running user application.

If the system administrator wants to have /var/scored/ directory in another file system, then the administrator must create a symbolic link before scored is run. In multi-user mode, checkpointed process images are also stored in /var/scored directory if a user program requests checkpointing. This directory is also used for storing some temporary file(s) created by the scatter program or user programs. Although these files will be removed when a user job terminates, the directory must be located in a file system where there is enough disk space for holding those files possibly created by multiple users simultaneously.

There can be multiple temporary file spaces in /var/scored/multiuser/. Each of them is called session. A session is identified by an associated session ID (number) and the SCore-D server hostname. At the SCore-D invokation, a session can be specified with -session and -server options. Be default, session number is set to zero (0). When SCore-D is invoked and the session having the same session number already exists on the same server host, but the cluster configuration is different, then SCore-D aborts. A session can be reset (cleared) when the -reset option, and the -session options if you want to specify a particular session, is specified, and all sessions can be reset by the -resetall option.

The options available for scored:

-ts time_slice
Specify the time slice (in milliseconds) for gang scheduling. The default is 500 msec

-syslog address[:port]
Specify the address (and port) of the scbcast server for the system logger. The default port is 9902

-sysmon address[:port]
Specify the address (and port) of the scbcast server for the system monitor. The default port is 9904

-server server_node
Specify the node (host) number of the SCore-D server host which accepts users login. If this option is not specified, the default node to use is the last node in the cluster, as determined by the order of nodes in the SCore Cluster Database, scorehosts.db(5)

-v
Print version information on standard output, then exit

-restart
When SCore-D is invoked with this option it tries to restart user processes which were terminated due to SCore-D crashing. When a user process has been checkpointed, the user process will be restarted from where the most recent checkpoint has been created.

-network network_name
Specify a network for SCore-D to communicate with in a cluster. Valid network names can be found in the scorehosts.db file. This network specification is independent from the network for user parallel processes to communicate.

-operator username
This option specifies a user to control SCore-D via sc_console(8) command as a system administrator. If this options is not specified, the username who invokes SCore-D is used.

-session session-no
Specify session number.

-reset
When SCore-D is invoked with this option, firstly it cleanes up the session status in the SCore-D directory, and the all checkpointed data in the session will be lost.

-resetall
When SCore-D is invoked with this option, firstly it cleanes up all files in the SCore-D directory, and the all checkpointed data of all session will be lost.

Files
/var/scored/
SCore-D creates a number of temprary files in this directory.

SEE ALSO
scout(1), scrun(1), scorehosts.db(5), scoreboard(8), sc_console(8) scbcast(8), SCore-D Job Scheduling Scheme, Automatic Operation and High Availability of SCore-D

$Id: scored.html,v 1.6 2003/03/26 01:25:59 kameyama Exp $