SUMMARY: CPU load on multiprocessor server

Sorry i took so long to get this summary on the list....

Anyways, big thanks to the following people for their responses:
	Darren Dunham
	Jay Lessert
	Steve Mickeler
	Matthew Stier
	Bill McCaffrey
	Riddoch, John
	Hendrik Visage
	Johan Hartzenberg
	Haywood, Steven
	Adam Fathauer
	Bob Rahe
	Wanke Matthias

The general concensus was that the load average was for all CPUs.
Basically that is the average number of processes waiting for te
entire system (independent of the number of CPUs in the system).
As a rule of thumb you could consider a load of 1 on a single CPU
server the same as a load of 4 on a 4 CPU server.

Also it was pointed out that the load average will mean different
things depening on what kind of work the server is doing.

From Adam Fathauer:
"...some single processor platforms that run with a sustained load of
4 that are still quite responsive and others that run under a load of
1 that are very busy."

From Wanke Matthias:
"A rule of thumb:
- on an Online Transaction System where response time ist important,
avoid a situation where idle time goes below 15% for more than 5
minutes.
- on number cruncher or datawarehouses, you can drive it to maximum
througput and bring idle time to under 1%. There shouldn't be any
backdraws in modern operating systems anymore with high loads. We have
a dwh under old 2.6/24 CPUs which we drove to the edge with a load of 800(!)
- things where slow, but went along fine!
You should always combine system utilization and queue length to
determine your systems "real load"."

peace,
dannyB
--------------------------
Daniel Baldonado


----- Original Message -----
> I was curious about the load average values given by "w" and
"uptime".
> >From the "w" manpage:
> "...the average number of jobs in the run queue over the last 1,
> 5 and 15 minutes."
>
> I thought this number was specific for EACH CPU in a multiprocessor
> machine.  For example, if the load given by "w" on a 4 CPU system
> shows an average of 3.20 then there are an average of 3.20 jobs
> waiting for each CPU.
>
> I've also heard that the load average number given is for ALL the
CPUs
> and to get the actual load you should divide by the number of CPUs
on
> that machine.  For example, if the load given by "w" on a 4 CPU
system
> shows an average of 3.20 then there is an average of 0.80 jobs
waiting
> for each CPU.
>
> Can someone clarify this for me?
>
> I have an E450 server with 4 CPUs with a load that averages anywhere
> from 3.0 - 7.0.  I was always of the understanding that a load of
> less than 1 was an underutilized server, a load between 1-2 was a
> busy server, and anything over 2 was an overloaded server.  What are
> other people's guidelines on this?
_______________________________________________
sunmanagers mailing list
sunmanagers@sunmanagers.org
http://www.sunmanagers.org/mailman/listinfo/sunmanagers