Those interested in testing the various efforts focused on improving Linux desktop performance may wish to try the patches available from the CPU Scheduler Evaluation project page. Peter Williams announced his latest patch against the 2.6.7 stable Linux kernel which provides the ability to select a CPU scheduler at runtime. The current version of the patch allows a user to select either Con Kolivas [interview]' staircase scheduler [story], or Peter's own priority based scheduler.

Other patchsets designed to improve performance and worth trying out include Con Kolivas' -ck patchset, currently at 2.6.7-ck3 [story], and Nick Piggins [interview]' -np patchset, currently at 2.6.7-np2 [story]. Con's patchset which applies against the plain 2.6.7 kernel is described as, "patches designed to improve system responsiveness with specific emphasis on the desktop, but has scheduler changes suitable/configurable to any workload". Nick's patchset, which applies against the 2.6.7-mm3 kernel, provides a different CPU scheduler and additional memory management work. For the interested, Nick advises, "if anyone is having swapping or interactivity problems, please try it out."

From: Peter Williams [email blocked]
To: Linux Kernel Mailing List [email blocked]
Subject: [PATCH] CPU scheduler evaluation tool
Date: 	Mon, 28 Jun 2004 16:37:33 +1000

To facilitate the comparative evaluation of alternative CPU schedulers a 
patch that allows the run time selection of CPU scheduler between 
version 7.7 of Con Kolivas's staircase scheduler ("sc") and the priority 
based scheduler with interactive and throughput bonuses ("pb") has been 
created.  This patch is available for download at:

<http://prdownloads.sourceforge.net/cpuse/patch-2.6.7-spa_hydra_FULL-v1.2?download&gt;

The file /proc/sys/kernel/cpusched/mode has been added to those provided 
by the "pb" to control which of the schedulers is in control.  The 
string "sc" is used to select the staircase scheduler and "pb" to select 
the priority based scheduler described above.  The staircase scheduler 
control parameters "compute" and "interactive" have been moved into 
/proc/sys/kernel/cpusched along with the "pb" scheduler control 
parameters.  The scheduler starts in the "sc" mode. A primitive 
Glade/PyGTK GUI that provides the ability to switch between schedulers 
and to control scheduler parameters is available at:

<http://prdownloads.sourceforge.net/cpuse/gcpuctl_hydra-1.0.tar.gz?download&gt;

This GUI should also work with a standard version of Con Kolivas's 
staircase scheduler as well as the version based on my single priority 
array patch:

<http://prdownloads.sourceforge.net/cpuse/patch-2.6.7-spa_sc_FULL-v1.2?download&gt;

and the basic priority based scheduler:

<http://prdownloads.sourceforge.net/cpuse/patch-2.6.7-spa_pb_FULL-v1.2?download&gt;

and, for those so inclined, these patches broken into smaller parts for 
easier digestion are available at:

<https://sourceforge.net/projects/cpuse/>

An entitlement based "eb" scheduler will be added to the selection 
available in the near future.

Controls:

base_promotion_interval -- (milliseconds) controls the interval between 
successive promotions (is multiplied by the number of active tasks on 
the CPU in question) NB no promotion occurs if there are less than 2 
active tasks

time_slice -- (milliseconds) the size of the time slice (i.e. how long 
it will be allowed to hold the CPU before it is kicked off to allow 
other tasks a chance to run) that is allocated to a task when it becomes 
active or finishes a time slice.  (min is 1 millisec and max is 1 second).

max_ia_bonus -- (a value between 0 and 10) that determines the maximum 
interactive bonus that a task can acquire

initial_ia_bonus -- (a value between 0 and 10) that determines the 
initial interactive bonus that a newly forked task will be given.  This 
value will be capped by the max_ia_bonus.

ia_threshold  -- (parts per thousand) is the sleep to (sleep + on_cpu) 
ratio above which a task will have its interactive bonus increased 
asymptotically towards the maximum

cpu_hog_threshold  -- (parts per thousand) is the usage rate above which 
a task will be considered a CPU hog and start to lose interactive bonus 
points if it has any

max_tpt_bonus -- (a value between 0 and 9) that determines the maximum 
throughput bonus that tasks may be awarded

log_at_exit - (0 or 1) turns off/on the logging of tasks' scheduling 
statistics at exit.  This feature is useful for determining the 
scheduling characteristics of relatively short lived tasks that run as 
part of some larger job such as a kernel build where trying to get time 
series data is impractical.

compute -- (0 or 1) turn on/off the staircase schedulers "compute" switch

interactive -- (0 or 1) turn on/off the staircase schedulers 
"interactive" mode

Peter
-- 
Peter Williams [email blocked]

"Learning, n. The kind of ignorance distinguishing the studious."
  -- Ambrose Bierce

• Archive of above thread