Load Averages on Multiprocessors

[Thomas R. Statnick]

In article <130401 at uunet.UU.NET> rbj at uunet.UU.NET (Root Boy Jim) writes:
>Should the average length of the run queue be divided by the
>number of processors in the system or reported as is? Why?

   Well, I know of several vendors that report the load average "as is"
(companies like Digital, Stardent, and Gould come to mind).  By definition,
the load average is defined as "an average of the number of processes ready
to run or waiting for disk I/O to complete, as sampled over the previous
1-minute interval of system operation" (at least that is the way one book
defines it!).

   With the previous in mind, then it really does not seem to make since
to divide by the number of CPUs that the system has, and still call it
a "true" load average.  If you divide by the number of CPUs, it is more
a measure of whether all of the CPUs are non-idle, right?  An example
of where the a-la Sequent thinking goes wrong is when you have, say in
a bad case, 10 jobs sitting in device wait for exactly a minute (yes,
I know that this is not really a practical example, but for example sake)..
On a 10 CPU Sequent, the load would be 1.00, whereas the number of processes
ready to run is really 10..  The number of CPUs does not change the number
of jobs in the run queue...

   To answer why, maybe someone from sequent.com will answer that one...
I think it is more for the effect of saying that "our machine is supporting
100 active users and only running a load average of 1!"...  Sorry, Charlie,
thats a load of 10 in my book...


    trs