illumos: manual page: FSS.4

FSS(4) Device and Network Interfaces FSS(4)

NAME

FSS - Fair share scheduler

DESCRIPTION

The fair share scheduler (FSS) guarantees application performance by
explicitly allocating shares of CPU resources to projects. A share
indicates a project's entitlement to available CPU resources. Because
shares are meaningful only in comparison with other project's shares, the
absolute quantity of shares is not important. Any number that is in
proportion with the desired CPU entitlement can be used.

The goals of the FSS scheduler differ from the traditional time-sharing
scheduling class (TS). In addition to scheduling individual LWPs, the FSS
scheduler schedules projects against each other, making it impossible for
any project to acquire more CPU cycles simply by running more processes
concurrently.

A project's entitlement is individually calculated by FSS independently for
each processor set if the project contains processes bound to them. If a
project is running on more than one processor set, it can have different
entitlements on every set. A project's entitlement is defined as a ratio
between the number of shares given to a project and the sum of shares of
all active projects running on the same processor set. An active project
is one that has at least one running or runnable process. Entitlements are
recomputed whenever any project becomes active or inactive, or whenever the
number of shares is changed.

Processor sets represent virtual machines in the FSS scheduling class and
processes are scheduled independently in each processor set. That is,
processes compete with each other only if they are running on the same
processor set. When a processor set is destroyed, all processes that were
bound to it are moved to the default processor set, which always exists.
Empty processor sets (that is, sets without processors in them) have no
impact on the FSS scheduler behavior.

If a processor set contains a mix of TS/IA and FSS processes, the fairness
of the FSS scheduling class can be compromised because these classes use
the same range of priorities. Fairness is most significantly affected if
processes running in the TS scheduling class are CPU-intensive and are
bound to processors within the processor set. As a result, you should
avoid having processes from TS/IA and FSS classes share the same processor
set. RT and FSS processes use disjoint priority ranges and therefore can
share processor sets.

As projects execute, their CPU usage is accumulated over time. The FSS
scheduler periodically decays CPU usages of every project by multiplying it
with a decay factor, ensuring that more recent CPU usage has greater weight
when taken into account for scheduling. The FSS scheduler continually
adjusts priorities of all processes to make each project's relative CPU
usage converge with its entitlement.

While FSS is designed to fairly allocate cycles over a long-term time
period, it is possible that projects will not receive their allocated
shares worth of CPU cycles due to uneven demand. This makes one-shot,
instantaneous analysis of FSS performance data unreliable.

Note that share is not the same as utilization. A project may be allocated
50% of the system, although on the average, it uses just 20%. Shares serve
to cap a project's CPU usage only when there is competition from other
projects running on the same processor set. When there is no competition,
utilization may be larger than entitlement based on shares. Allocating a
small share to a busy project slows it down but does not prevent it from
completing its work if the system is not saturated.

The configuration of CPU shares is managed by the name server as a property
of the project(5) database. In the following example, an entry in the
/etc/project file sets the number of shares for project x-files to 10:

x-files:100::::project.cpu-shares=(privileged,10,none)

Projects with undefined number of shares are given one share each. This
means that such projects are treated with equal importance. Projects with
0 shares only run when there are no projects with non-zero shares competing
for the same processor set. The maximum number of shares that can be
assigned to one project is 65535.

You can use the prctl(1) command to determine the current share assignment
for a given project:

$ prctl -n project.cpu-shares -i project x-files

or to change the amount of shares if you have root privileges:

# prctl -r -n project.cpu-shares -v 5 -i project x-files

See the prctl(1) man page for additional information on how to modify and
examine resource controls associated with active processes, tasks, or
projects on the system. See resource_controls(7) for a description of the
resource controls supported in the current release of the Solaris operating
system.

By default, project system (project ID 0) includes all system daemons
started by initialization scripts and has an "unlimited" amount of shares.
That is, it is always scheduled first no matter how many shares are given
to other projects.

The following command sets FSS as the default scheduler for the system:

# dispadmin -d FSS

This change will take effect on the next reboot. Alternatively, you can
move processes from the time-share scheduling class (as well as the special
case of init) into the FSS class without changing your default scheduling
class and rebooting by becoming root, and then using the priocntl(1)
command, as shown in the following example:

# priocntl -s -c FSS -i class TS
# priocntl -s -c FSS -i pid 1

CONFIGURING SCHEDULER WITH DISPADMIN

You can use the dispadmin(8) command to examine and tune the FSS
scheduler's time quantum value. Time quantum is the amount of time that a
thread is allowed to run before it must relinquish the processor. The
following example dumps the current time quantum for the fair share
scheduler:

$ dispadmin -g -c FSS
#
# Fair Share Scheduler Configuration
#
RES=1000
#
# Time Quantum
#
QUANTUM=110

The value of the QUANTUM represents some fraction of a second with the
fractional value determined by the reciprocal value of RES. With the
default value of RES = 1000, the reciprocal of 1000 is .001, or
milliseconds. Thus, by default, the QUANTUM value represents the time
quantum in milliseconds.

If you change the RES value using dispadmin(8) with the -r option, you also
change the QUANTUM value. For example, instead of quantum of 110 with RES
of 1000, a quantum of 11 with a RES of 100 results. The fractional unit is
different while the amount of time is the same.

You can use the -s option to change the time quantum value. Note that such
changes are not preserved across reboot. Please refer to the dispadmin(8)
man page for additional information.

NAME

DESCRIPTION

CONFIGURING SCHEDULER WITH DISPADMIN

SEE ALSO