Charting server load
Lennart Sorensen
lsorense-1wCw9BSqJbv44Nm34jS7GywD8/FfD2ys at public.gmane.org
Thu Feb 1 21:40:06 UTC 2007
On Thu, Feb 01, 2007 at 04:03:31PM -0500, Ian Petersen wrote:
> I can see how a system with high load might still be responsive
> because I would define responsive as "responds nearly immediately to
> user input" where "user input" could be defined as key presses, mouse
> movements, ssh connection requests, etc. This could easily be
> accomplished by running user-facing processes at high priority so that
> user input forces the relevant processes onto the CPU.
>
> On the other hand, a non-hyperthreaded, single-core uniprocessor can
> still only do "one thing at a time", ignoring vector instructions, so
> if, for the last minute, there was an average of 50 ready processes in
> the run-queue, how is that not "overloaded"? A load of 50, to me,
> means that at any given moment the kernel had 50 things to do but had
> to pick one (or N on an N-core system). If I had 50 things to do at
> any given moment and was restricted to doing only one at a time, I'd
> feel pretty stressed.
If the 50 processes each needed 2.01% of the cpu to get all their work
done all the time, then you are only overloaded by 0.5%, so really you
are keeping up with the load. If even one of the processes was to be
stopped, your load should actually drop quickly to less than 1 since it
would suddenly be able to keep up with all the work. Load averages are
rather weird that way. Another system could have 2 tasks running which
both need all the cpu they can get, which would mean by some definitions
that the system is overloaded by 100%, although of course if they are
really cpu bound no matter how fast your cpu, they would still want 100%
of it until the task is done. You could have 2 tasks, each of which need
75% of the cpu to get all their work done real time, but by having two,
you get a load average of 2 and both tasks are not quite keeping up with
their work in real time.
--
Len Sorensen
--
The Toronto Linux Users Group. Meetings: http://gtalug.org/
TLUG requests: Linux topics, No HTML, wrap text below 80 columns
How to UNSUBSCRIBE: http://gtalug.org/wiki/Mailing_lists
More information about the Legacy
mailing list