/ 
Schedule-Load-1.5
/ Schedule::Load

NAME

Schedule::Load - Load distribution and status across multiple host machines

SYNOPSIS

# Get per-host or per top process information
use Schedule::Load::Hosts;
foreach my $host ($hosts->hosts) {
    printf $host->hostname " is on our network\n";
}

# Choose hosts
use Schedule::Load::Schedule;
my $scheduler = Schedule::Load::Schedule->fetch();
print "Best host for a new job: ", $scheduler->best();

# user access
rtop
rschedule reserve <hostname>

DESCRIPTION

This package provides useful utilities for load distribution and status across multiple machines in a network. To just see what is up in the network, see the rschedule or rtop, rloads or rhosts commands.

The system is composed of four unix programs (each also with a underlying Perl module):

slreportd

slreportd is run on every host in the network, usually started with a init.d script. It reports itself to the slchoosed daemon periodically, and is responsible for checking loading and top processes specific to the host that it runs on.

slreportd may also be invoked with some variables set. This allows static host information, such as class settings to be passed to applications.

slchoosed

slchoosed is run on one host in the network. It collects connections from the slreportd reporters, and maintains a internal database of the entire network. User clients also connect to the chooser, which then gets updated information from the reporters, and returns the information to the user client. As the chooser has the entire network state, it can also choose the best host across all CPUs in the network.

rschedule

rschedule is a command line interface to this package. It and the aliases rtop, rhosts, and rloads report the current state of the network including hosts and top loading. rschedule also allows reserving hosts and setting the classes of the machines, as described later.

slpolice

slpolice is a optional client daemon which is run as a cron job. When a user process has over a hour of CPU time, it nices that process and sends mail to the user. It is intended as a example which can be used directly or changed to suit the system manager preferences.

MODULES

Schedule::Load::Hosts

Schedule::Load::Hosts provides the connectivity to the slchoosed daemon, and accessors to load and modify that information.

Schedule::Load::Schedule

Schedule::Load::Schedule provides functions to choose the best host for a new job, reserving hosts, and for setting what hosts specific classes of jobs can run on.

Schedule::Load::Reporter

Schedule::Load::Reporter implements the internals of slreportd.

Schedule::Load::Chooser

Schedule::Load::Chooser implements the internals of slchoosed.

RESERVATIONS

Occasionally clusters have members that are only to be used by specific people, and not for general use. A host may be reserved with rschedule reserve. This will place a special comment on the machine that rschedule hosts will show. Reservations also prevent the Schedule::Load::Schedule package from picking that host as the best host.

To be able to reserve a host, the reservable variable must be set on that host. This is generally done when slreportd is invoked on the reservable host by using slreportd reservable=1.

CLASSES

Different hosts often have different properties, and jobs need to be able select a host with certain properties, such as hardware or licensing requirements. Classes are generally just boolean variables which start with class_. Classes can be specified when slreportd is invoked on the slreportd class_foo=1. The class setting may be seen with rschedule classes or may be read (as may any other variable) as a accessor from a Schedule::Load::Hosts::Host object.

Once a class is defined, a scheduling call can include it the classes array that is passed when the best host is requested. Only machines which match one of those classes will be selected.

COMMAND COMMENTS

rschedule loads or rloads show the command that is being run. By default this is the basename of the command invoked, as reported by the operating system. Often this is of little use, especially when the same program is used by many people. The rschedule cmnd_comment command or Schedule::Load::Schedule::cmnd_comment function will assign a more verbose command to that process id. For example, we use dc_shell, and put the name of the module being compiled into the comment, so rather then several copies of the generic "dc_shell" we see "dc module", "dc module2", etc.

HOLD KEYS

When a best host is picked for a new job, there is often a lag before a process actually starts up on the selected host, and enough CPU time elapses for that new process to claim CPU time. To prevent another job from scheduling onto that host during this lag, scheduling calls may specify a hold key. For a limited time, the load on the host will be incremented. When the job begins and a little CPU time has elapsed a hold_release call may be made, (or the timer expires), which releases the hold. This will cause the load reported by rschedule hosts to occasionally be higher then the number of jobs on that host.

FIXED LOADS

Some jobs have CPU usage patterns which contain long periods of low CPU activity, such as when doing disk IO. make is a typical example; the parent make process uses little CPU time, but the children of the make pop in and out of the cpu run list.

When scheduling, it is useful to have such jobs always count as one (or more) job, so that the idle time is not misinterpreted and another job scheduled onto that machine. Fixed loading allows all children of a given parent to count as a given fixed CPU load. Using make again, if the parent make process is set as a fixed_load of one, the make and all children will always count as one load, even if not consuming CPU resources. The rschedule loads or rloads command includes not only top cpu users, but also all fixed loads. If a child process is using CPU time, that is what is displayed. If no children are using appreciable CPU time (~2%), the parent is the one shown in the loads list.

DISTRIBUTION

The latest version is available from CPAN.

SEE ALSO

User program for viewing loading, etc:

rschedule

Daemons:

slreportd, slchoosed, slpolice

Perl modules:

Schedule::Load::Chooser, Schedule::Load::Hosts::Host, Schedule::Load::Hosts::Proc, Schedule::Load::Hosts, Schedule::Load::Reporter, Schedule::Load::Schedule

AUTHORS

Wilson Snyder <wsnyder@wsnyder.org>