User + sys times greater than elapsed time

Thread view:

Enable EMail Alerts

Start New Thread

Thread rating:

mweallans@panacea.co.uk - 30 Sep 2005 18:29 GMT

Previously in these hallowed halls I have queried the values for user
and sys cpu as reported by sysmaster. The documentation says that it
reports minutes but I have proved that it is seconds.

The today I hit another "strange effect". I have been trying to look
into a performance problem so I have written a script which basically
selects user and os times from the sysmaster database and stores them
with the current time.

I run this script at 10 minute intervals throughout the day. Imagine
my surprise when I find that user time + sys time comes out to greater
than elapsed time * number of processors.

Does anybody have any idea why?

regards

Malcolm

Reply to this Message

Madison Pruet - 30 Sep 2005 20:05 GMT

Because the times are measured in units of CPU clock ticks, not as a
continuum ;-)

Here's the problem. If the length of the clock tick is measured as a finite
unit (which it is), then how can you split that unit? (you can't). So if
the control changes from a user control to the kernel control within a clock
tick as opposed to the cusp of the clock tick, then the time measured is
going to be a bit fuzzy because both the system and the user will be granted
that clock unit, eventhough the system only used a percentage of the clock
tick and the user used the other part.

IDS gets it's breakdown from the same place that timex.

> Previously in these hallowed halls I have queried the values for user
> and sys cpu as reported by sysmaster. The documentation says that it
[quoted text clipped - 14 lines]
>
> Malcolm

Reply to this Message