top problem - Load averages

[ericfjchen]

How to judge the CPU performance via top--load averages? Does any document mention it? Pls share your experience.

Thanks for any comment

FJ

[Eric Antunes]

Hi Eric,

My experience tells me that CONSISTENT loads above 10 are bad. Nevertheless, you can have spikes above those 10 but it is bad is the system stays there.

Best Regards,

Eric Antunes

[Bill Hassell]

The load average doesn't mean what it sounds like. The load average is the average number of processes that are running or waiting to run (the kernel's runqueue). In a one-processor system, a load average of 2 or 3 is fine or it could be crawling...it all depends on what the processes are doing. If you have 16 processors, then a load of 10 is nothing.

Generally, loads that cause slowdowns for everyone are related to disk activity rather than CPU usage.

A very high CPU usage seldom causes any apparent slowdown for users. That's because after a few seconds of 100% CPU usage, the priority of these programs is adjusted to allow other short-lived programs to run as needed. You can use 100% CPU with the simple set of commands:

while :
do
:
done

This starts the shell into an infinite loop (CTRL-C to terminate) which consumes 100% of one processor. Start dozens or hundreds of these processses and you'll still be able to login and run vi, etc. This is the nature of timesharing.

Occasionally, a runaway process or a bug in the kernel may cause runaway system activity such that the load factor is very high (5-10 times the number of porcessors) and this would be fixed by tracking down the process or installing the latest patches.

[Eric Antunes]

Hi,

In my previous comment, I was talking of my experience in my systems and those are one-processor ones...

Best Regards,

Eric Antunes

[Steven E. Protter]

Do you have any other symptons other than load average. I usually don't act until there is a complaint or noticable issues. That doesn't mean I don't look, but top never got me excited.

top is the least accurate of available tools. Does glance/gpm or sar show the same issue?

SEP

[C. Beerse_1]

My experience with load-averages and cpu-s is as follows:

As long as the load is below the number of cpu-s, you will not get a much better performance if you get more cpu-s (performance boost will be minimal).

Once the load is more than the number of cpu-s, you will get some performance improvements if you install/activate more cpu-s.

My experience: I had more cpu-s than load and installed seti@home. I saw the load increase by (about) 1 for each seti@home instance.