[CoLoCo] Odd load for sleep mode

[Scott Scriven]

* Chris Yoder <chris at medioh.com> wrote:
> Or you could just run top in a terminal window, but it isn't as pretty.

Top is okay, but it misses some important resources, and the 
display isn't as human-readable as a graph.

> I'd have to read the source to make sure, but normally a 
> process blocked on IO does not contribute to the load,  because 
> the load average is the number of threads waiting to get onto 
> the cpu, as a simplistic rule of thumb.

Nope, the load average isn't just CPU time.  It's the number of 
processes waiting on the kernel for any reason at all.  For a 
demonstration, try this...  Open two shells.  In each, run 'cat 
/dev/sda > /dev/null'.  Let it run for a while, and watch the 
system load and the CPU use.  The loadavg should stabilize right 
at 2, while the CPU load stays down around 10%.

Another easy way to show it is to mount a dir via NFS, unplug the 
network, then try to 'ls' the dir.  Do this a dozen times, and 
the load will rise to 12, but the system won't actually be doing 
anything -- no CPU activity, no disk activity, and barely any 
network activity.

An odd quirk is that you can actually have a load of 0.00 with 
90% CPU activity.  To do this is a little trickier, but fairly 
simple.  Just make a program like this:

  while(1) {
    for(i=0; i<1000000; i++) {} // requires calibration
    usleep(1);  // 1 microsecond
  }

The idea is to eat up most of a time slice, then voluntarily give 
control back to the kernel.  Even though it's using most of the 
CPU capacity, the process won't contribute to the load average at 
all.  The kernel calculates the load average at the end of each 
time slice, at which point the task isn't waiting for anything 
from the kernel.

That effect isn't usually so noticable, but I often see a load of 
0.00 on an "idle" system while the CPU is 5%-20% busy.

Anyway, all I was getting at is that I don't often find loadavg 
useful.  It's a data point, but not one I place much value on.


-- Scott