[LAU] getting cpu usage on commandline

[Cedric Roux]

----- "Atte André Jensen" <atte.jensen at gmail.com> wrote:

> Julien Claassen wrote:
> > Hello atte!
> >   Two things I can think of: Have you tried the normal top? - I
> don't 
> > know about its accuracy.
> 
> I played around a bit more, and it seems I misjudged "ps aux", since
> it 
> seems to report correct values. Will takt that route and get back in 
> case of further trouble.
> 
> Thanks for the input, anyhow!

Short solution: top -b -n 1
(with procps 3.2.7, just in case it
fails for you...)

Long explaination (you may skip it):
Note that ps results are not accurate.
The CPU usage is reported as % of time spent
running during the entire lifetime of a process.
So if you have a process sleeping for one minute
and then running like hell for one second, you might
think it won't eat much CPU if you ps it during
this second even if in reality it eats up all
the CPU.

The /proc/[processID]/stat gives you statistics
on the process. You are interested in the field
utime I guess, which gives you the number of
jiffies the process smoked in user mode at the
time of reading /proc/[processID]/stat. (The
output format of /proc/[processID]/stat is
described in Documentation/fs/proc.txt found in
the linux sources tarball.)

So to know how much CPU the process smokes you
first read /proc/.../stat, wait a bit (1s or so),
read /proc/.../stat again and do something
like (value2 - value1) / nb_jiffies_of_1s
(nb_jiffies_of_1s might be defined on your
system in /usr/include/asm/param.h, it's the
HZ value. It might also exist in your shell's
environment (try: echo $HZ). Or you can go
read the procps sources (in procps-3.2.7
this stuff is in proc/sysinfo.c, and it is a bit
tricky to get this value it seems...)
to see how they get it.
You can also do:
cat /proc/stat; sleep 1; cat /proc/stat
and see how bigger the 4th number on
the line "cpu: XX" is.
(Or /proc/self/stat and the 22nd number.)
On my host HZ=100, but on yours it
might well be 250, 300 or 1000 (this
are the values I see when doing
a "make config" of the kernel).
I guess 1000 is the most probable.

Anyway, there is no way to know how much CPU a given
process is eating at a given moment. You have to
"wait a bit" and see what changed.

But maybe you are more interested in the
global times? in which case just do
"time command" and you get some stats
at the end, when the process exits.

For the "accurate" way to measure
it's not too hard to code. The program would
open /proc/[processID]/stat, get the number,
wait, open again, output the result and exit.
A few lines of C code.

Or you can do "top -b -n 1" with other options
if you like. Just read the manpage.

(Note that it may not be very accurate either
if the process switches before the end of
a jiffy for example; I had such a problem in the
past with a 2.0 something kernel; the process
was reported as eating 0% CPU which was
false. But maybe things are different
now.)

Hope it helps.
Cédric.