Load Average in a multi CPU system

[Amit Manna_6]

Can anybody tell me what does the load average signify and how its calculated and whats the unit of measurment for the same

I am attaching the o/p here.

[Steven E. Protter]

A thread:
http://www1.itrc.hp.com/service/james/dispDoc.do?docURL=http%3A%2F%2Fsearch.hp.com%2Fredirect.html%3Furl%3Dhttp%253A%2F%2Fforums1.itrc.hp.com%2Fservice%2Fforums%2Fquestionanswer.do%253FthreadId%253D852434%26qt%3D%252BLoad%2B%252BAverage%2B%26hit%3D1&aid=SEARCH_FORUMS&pil=1&serStr=Load+Average&pir=1

Educational doc:
http://www1.itrc.hp.com/service/james/dispDoc.do?docURL=%2Fservice%2Fcki%2FdocDisplay.do%3FdocLocale%3Den_US%26docId%3D200000071337163&aid=SEARCH_CKI&pil=5&serStr=Load+Average

Explanation attempt:

Load Average is just that, an average load factor over the time you are running the utility, in your case, sar.

Load has to do with the number of processes waiting for CPU and not getting it over a certain period of time, measured in milliseconds.

If Bill Hassell sees this and responds, he may share a story about how a system he worked on had a load average of 240 and excellent response time.

Load average being high is sometimes an indicator of trouble. Sometimes not. Your system looks pretty busy.

Does it have any specific performance issues?

SEP

[Jeff Schussele]

Hi Amit,

The load average is simply a weighted average of processes in the run queue - i.e. processes *waiting* for the next available CPU cycles.
This means that the more processes waiting to run the more the system is "loaded".

HTH,
Jeff

[renarios]

Hi Amit,

Maybe obsolete to say, but another way to watch your system load is top or glance (mwa)

Cheerio,

Renarios

[Alberto Ochoa]

To add some info.
As per your posted doc. we can see the Load averages (the average number of processes the were active in the last one, five, and fifteen minutes (1.72, 1.69, 1.56 respectivelly)) of all the active processors in the system.
You can see that in the last minute the value is 1.72, and if you see in the
bottom of the column CPU, there is the AVG value for all your CPus which corresponds to 1.72, which is the sum of all the values under the column LOAD of your CPU's divided by your 8 CPU.

In this case it seems that for the last 15 minutes the system has had consistent load,
although it is being a bit increased as per the last minute value (1.72).
In a very general way a light load may be considered below 3, medium grade loads could be considered around 5~7, and loads above 8 could be considered heavy loads, but, this is just a general point of view. You'll see many opinions on this issue.
It would be good idea to monitor your system at a regular intervals to
see what is normal to your system.

[Bill Hassell]

The "load average" is the average value of the run_queue, nothing more. The run_queue is all programs that are currently running (using CPU) or ready to run but there is no free CPU. So a load factor of 9 means that 9 processes are running or waiting to run. It is somewhat related to CPU usage but not in the way the name "load" is used. You can create a load factor of 1 by running this simple script:

while :
do
:
done

This script will keep a CPU 100% busy. On a one-CPU system, starting 5 copies doesn't change the CPU usage (still 100%) but there are 4 copies waiting for a timeslice, so the load factor is 5. Is the one-processor system any more loaded? No. HP-UX simply switches between each script every 100ms (kernel value is timeslice, default is 10, measure is 10ms).

So, it is simply the number of processes running or waiting to run. If you have a 2 processor system, a run_queue (load factor) of 2 means both CPUs are working as fast as they can on the user programs. A load factor of 240 for a 2 processor system is quite easy to create (just run the above script 240 times). But since all the scripts are CPU intensive, interrupts such as logins and other disk access will take priority and the system will seem fairly normal.

In the real world of databases and design programs, the load isn't nearly as simple because programs run lots of instructions as well as I/O. So the metric is slightly useful but should not be a major part of preformance monitoring.