[ale] questions about load averages

Jason Boyles jason at alltel.net
Tue Nov 18 17:05:45 EST 1997


On Tue, 18 Nov 1997, Alexander Barton wrote:

> > [...]  On a single CPU machine a load average of one means
> > that the CPU is saturated. [...]
> (I may be wrong, but....)
> A load average of one doesn't mean the CPU has to be saturated.  You could
> have a load average of >1 and still have less than 100% saturation.

	True. A load average of N indicates that over the past minute at
each clock interrupt, on average, N process(es) was/were in the ready to
run state. "On average" is, of course, an important qualification.
	You're right, we should make the distinction between load average
and instantaneous load. Load average is a rough indication of CPU
utilization at best.

	Bursty loads where the load average is < 1 can be a problem for
those cases where low latency is important. I guess at that point you
should either be considering a real time OS, or implementing your task in
the kernel where you can disable interrupts when needed (in some cases
even that may not be good enough). Are there any ALE list members using RT
Linux? Any comments on its current state?

	The answer to the question of what a 'bad' load average is varies
according to the particulars of the environment. 

> Just for fun, I started four finds, plus top, plus everything else
> already on my desktop.  find (over NFS) is very not-CPU-intensive.
> I got a load average fluctuating around 1.74 while saturation
> was around 35% (100% - idle%).
> [ . . . ] Occasionally, this info arrived but had to wait
> because one or more finds were already using the CPU.  But, for most
> of the time (65% idle), all four finds and top and everything else
> was waiting on I/O.

	Load average can be deceptive. Something like vmstat or top can be
your friend in such circumstances.

> -Alexander






More information about the Ale mailing list