[wellylug] Load average statistics dubious

[Bruce Hoult]

On Tue, Mar 30, 2010 at 12:28 PM, Daniel Pittman <daniel at rimspace.net> wrote:
> Bruce Hoult <bruce at hoult.org> writes:
>> On Tue, Mar 30, 2010 at 12:29 PM, C.T.F. Jansen
>> <frank.jansen at actrix.gen.nz> wrote:
>>
>>> I used to use a number of different unix OS's and the load average
>>> statistics on them is an artificial number and often with values too
>>> different with what is actually happening on the system. It was regarded as
>>> a joke and people warned not to use it.
>>
>> I don't understand. Load average has a very specific definition -- the
>> number of processes in runnable (including actually running) state.
>
> It is worth noting that this may, or may not, include processes blocked
> waiting on I/O.  On some platforms like Linux this has even changed over time,
> since they now report with, but used to report without.

This is true.  In recent times the primary cause I've seen of load
average heading for the sky (dozens) is too many apache or mailscanner
or something instances running the thing out of RAM and causing
massive swapping.  More RAM is the right answer, but if you're lucky
then limiting the number of instances may stop the thrashing and
increase the throughput and suddenly the load average is under 1.0. (x
the number of CPUs...)