description | CPU load monitor |
owner | clamky@hotmail.com |
last change | Sun, 27 Jul 2014 17:15:12 +0000 (27 21:15 +0400) |
URL | git://repo.or.cz/apc.git |
| https://repo.or.cz/apc.git |
push URL | ssh://repo.or.cz/apc.git |
| https://repo.or.cz/apc.git (learn more) |
bundle info | apc.git downloadable bundles |
content tags
|
|
README
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
WARNING
The kernel module part of this program messes with internal affairs of
the kernel, while best effort was put into making it safe, there are:
NO GUARANTEES WHATSOEVER
Furthermore removing the previous versions of the module (via
rmmod(8)) caused one particular kernel version to panic
(2.6.8-2-686-SMP form Debian), to the best of my current knowledge
panics are only possible on SMP machines (and with maxcpus > 1). Pair
of safety nets were added and this particular kernel no longer panics
upon module removal, but, again, three words in caps above apply.
The module expects certain things not to happen at particular point in
execution, otherwise the information kernel module exports can not be
trusted. Those `things' did happen on aforementioned Debian kernel,
so if you need to run APC there you might want to uncomment first line
of `mod/itc-mod.c'.
2.4 series of kernels were never tested on SMP as such the module will
refuse to build for them.
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
This is APC - graphical CPU load meter.
It is more suitable/accurate in situations where applications generate
"short" periodic bursts of activity.
It works by measuring the time spent in the kernels idle handler. CPU
load time is taken to mean:
time spent in idle handler
1 - --------------------------
total time elapsed
Con Kolivas in his post on LKML (http://lkml.org/lkml/2007/2/12/7)
described the way Linux gathers information that it exports to
`/proc/stat' (at least for "boring" architectures), this method is by
no means accurate and can "lie" in either direction.
You can witness this by running the `hog' example and, if stars are
aligned correctly, you will notice that something is wrong with what
`/proc/stat' claims. Since most of the CPU monitoring applications use
`/proc/stat' they will produce incorrect results too.
Note that dyntick kernels will most likely produce different but
still incorrect accounting.
(For more information: http://lkml.org/lkml/2007/6/14/301 and bellow)
Kernel (starting with version 2.6.21) comes with a document describing
the way accounting is currently done and problems with this approach
(Documentation/cpu-load.txt)
Following thread describes a take on addressing the issue properly:
http://marc.info/?t=117480935100001&r=1&w=2
Apart from being inaccurate, `/proc/stat' exports monotonically
increasing load times but _NOT_ real time[1], so there's omni-present
sub-jiffy error. Not to mention that jiffy resolution is somewhat low.
If you are not running RT kernel and need sorta-kinda, semi-correct
load meter in situation when `/proc/stat' is disconnected with reality
APC might present a better choice.
The kernel module part of APC tries to insert itself as a power
management idle handler and when invoked measure how much time is
spent executing previous/default one - this information is represented
by yellow color, values obtained via `/proc/stat' are represented by
red.
You can use `-help' command line option to get a brief overview of
tunable parameters.
Tested on[2]:
Linux 2.4.30 - AMD Athlon(tm) Processor (1.4 Ghz)
Linux 2.6.17.6
Linux 2.6.19.2 - AMD Athlon(tm)64 X2 Dual Core Processor 3800+
Linux 2.6.18 - AMD Athlon(tm)64 3800+
Linux 2.6.18.3 - PowerPC 7447A
Linux 2.6.20.1 - PowerPC 7447A
Linux 2.6.19 - [some Core 2 Duo]
It's possible that RMClock[3] does something similar(load measuring
wise) on Microsoft Windows.
[1] Unlike `/proc/uptime'. But this one is useless for SMP
[2] SMP not tested on 2.4 kernels, nor QUIRK mode. SMP on PPC wasn't
tested either
[3] http://cpu.rightmark.org/products/rmclock.shtml
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
To build idlestat (boring console application) you will need:
GCC - http://gcc.gnu.org/
Plus all what is required to build a kernel module.
Process:
$ gcc -o idlestat idlestat.c
$ cd mod && make
Idlestat (as well as APC) requires kernel module to be loaded in order
for it to operate. Module loading is described below.
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
To build APC (graphical application with bells etc) you will need:
OCaml - http://caml.inria.fr/ocaml/
LablGL - http://wwwfun.kurims.kyoto-u.ac.jp/soft/lsl/lablgl.html
(and by extension some OpenGL implementation)
GLUT - http://www.opengl.org/resources/libraries/glut/
http://freeglut.sourceforge.net/
GCC - http://gcc.gnu.org/
Plus all what is required to build a kernel module.
Process:
<untar and go to directory with sources>
$ sh build.linux
# if following step fails(on X86) read next section before trying to
# execute mknod
$ su -c 'insmod mod/itc.ko' - 2.6 Kernels
$ su -c 'insmod mod/itc.o' - 2.4 Kernels
$ su -c "chmod +r /dev/itc"
[make sure you are in X]
$ ./apc
``````````````````````````````````````````````````````````````````````
Following applies only to Linux running on X86.
If the module fails to load consult dmesg(8). Most likely cause is the
lack of exported `default_idle' function. Few workarounds follow.
Workarounds
Variant 1
Add `idle=halt' to the kernel command line (method depends on
the boot-loader) and reboot.
Variant 2 (DANGEROUS)
--------------------------------------------------------------
Kernel 2.6
$ func=$(awk '/default_idle$/ {print "0x" $1}' /proc/kallsyms)
$ su -c "/sbin/insmod mod/itc.ko idle_func=$func"
--------------------------------------------------------------
Kernel 2.4
$ func=$(awk '/default_idle$/ {print "0x" $1}' /proc/ksyms)
$ su -c "/sbin/insmod mod/itc.o idle_func=$func"
======================================================================