[SOLVED] - trying to make sense of system load

Roken · 2012-07-03 09:47:57

I've already done my reading on this, and the matter should be fairly simple, so here's my dilemma.

For real time monitoring of system load I have a conky displaying:

CPU frequency and usage (per core) - taken from cpufreq and conky built in function
1minute, 5 minute, and 15 minute load averages - taken from uptime
Top 4 tasks(including individual CPU load for tasks) in Solaris mode (i.e. - average across 4 cores) - from top

The top information shows that my top 4 tasks rarely consume more than about 12% total CPU (usually 9% - 12%). However, the per core CPU usage shows total CPU usage in the region of 25-30% per core at any one time, usually at the lowest clock frequency (total of about 25-30% average across the system - frequency ranges in 4 steps from 800Mhz - 3.4Ghz depending on load)
Load averages report, usually, between 1.7 and 2.5

Now, I'm getting on a bit, and my maths may not be what it was, but aren't these four separate measures telling me wildly different things, and how do I reconcile everything I'm seeing so that it makes sense? By my reckoning I'm seeing a concurrent load report of 12%, 30% and 65% from the three measures.(And the 30% measure is questionable since it's 30% of underclocked power).

Processes outside of the top 4 reported by top typically consume 0.2 - 0.1 or 0%

Last edited by Roken (2012-07-03 20:18:06)

skunktrader · 2012-07-03 10:14:29

Some of us have had problems with load averages using recent kernels
https://bbs.archlinux.org/viewtopic.php?id=141289

Roken · 2012-07-03 10:24:05

yes, I saw that thread, but I'm not particularly concerned with load averages at the moment (given that I don't actually know what my real load average is). I have a server constantly running in a detached screen session, and so I would expect my load average to be higher than the norm. What I am concerned about is the apparent inconsistency between what the different tools are telling me.

Once I know what my "real" load average is, then I may share your other concerns, but not just yet.

donniezazen · 2012-07-03 16:14:15

https://bugs.archlinux.org/task/29850#comment95791

The thread that I started.

I don't know how to compile a kernel, yet. I am using linux-ice from AUR which 3.2 and is very power friendly.

Last edited by donniezazen (2012-07-03 16:15:54)

karol · 2012-07-03 17:55:42

donniezazen wrote:

I don't know how to compile a kernel, yet. I am using linux-ice from AUR

Does not compute ;P

donniezazen · 2012-07-03 18:13:18

@Karol I don't know what you mean.

alphaniner · 2012-07-03 18:29:42

donniezazen wrote:

I don't know what you mean.

If you're running a kernel from the AUR, then you compiled it.

Pres · 2012-07-03 18:40:15

http://www.teamquest.com/resources/gunt … /index.htm

That article might be of interest to you. The important take away is that load averages are not based on cpu utilization, but run queue length.

Here's another article that tries to simplify it:
http://blog.scoutapp.com/articles/2009/ … d-averages

donniezazen · 2012-07-03 19:33:01

@alphaniner That is correct. I wanted to use 3.3. But at the moment I don't have enough time to look into kernel compilation. Installing(compiling) from AUR is very straight forward.

Roken · 2012-07-03 20:17:47

donniezazen wrote:

https://bugs.archlinux.org/task/29850#comment95791The thread that I started.

Thank you. It wasn't quite what I was looking for, but may solve why I've set up such monitoring in the first place. Installing linux-ice now to test.

@Pres

Thank you so much. That helps clear some things up, and means I can remove the load averages from my monitoring (they aren't measuring what I want to measure).

I still need to figure out the disparity between per process CPU usage and overall CPU usage, but those two were closer together anyway, so less of a worry.

I'm going to mark this thread solved.

Roken · 2012-07-03 23:26:29

donniezazen wrote:

https://bugs.archlinux.org/task/29850#comment95791The thread that I started.

I've confirmed this as an upstream bug in 3.4 (not a ck issue). load average tells me a lot (currently 1 minute load average is 0.15 on 3.2 kernel compared with 1.9 in the same situation with 3.4, and I've dropped 9 degrees C between the 2)

I'm assuming that you've already reported it. If you've been given any indication of what causes it (kernel config) I'll happily try a rebuild of 3.4 to see if it fixes.

Arch Linux

#1 2012-07-03 09:47:57

[SOLVED] - trying to make sense of system load

#2 2012-07-03 10:14:29

Re: [SOLVED] - trying to make sense of system load

#3 2012-07-03 10:24:05

Re: [SOLVED] - trying to make sense of system load

#4 2012-07-03 16:14:15

Re: [SOLVED] - trying to make sense of system load

#5 2012-07-03 17:55:42

Re: [SOLVED] - trying to make sense of system load

#6 2012-07-03 18:13:18

Re: [SOLVED] - trying to make sense of system load

#7 2012-07-03 18:29:42

Re: [SOLVED] - trying to make sense of system load

#8 2012-07-03 18:40:15

Re: [SOLVED] - trying to make sense of system load

#9 2012-07-03 19:33:01

Re: [SOLVED] - trying to make sense of system load

#10 2012-07-03 20:17:47

Re: [SOLVED] - trying to make sense of system load

#11 2012-07-03 23:26:29

Re: [SOLVED] - trying to make sense of system load

Board footer