--------- Forwarded Message

Date:    Mon, 03 Mar 1997 10:53:23 +0000
From:    Patrick Sansom <[EMAIL PROTECTED]>
To:      Denis Howe <[EMAIL PROTECTED]>
cc:      Patrick Sansom <[EMAIL PROTECTED]>
Subject: Re: Measuring CPU time 


> From: Denis Howe <[EMAIL PROTECTED]>
> Date: Fri, 28 Feb 97 23:50:56 GMT
> > Comparing user times may give unrealistic comparisons on very
> > heavily loaded machines.
> 
> Why?

In my experiaence, heavy paging does affect the user time since there is
some user time spent processing / restarting from page faults --- running
with more memory can reduce user time. 

If the paging is caused by other sources of load on the machine this
introduces just another source of variation into the measurements. If
all runs are under the same conditions then the relative measurements
should not change -- but how do we know the same conditions are`
prevailing. Similarly, frequent context swithes affect cache behaviour.

These are not crucial when getting a ball park figure/comparison, but 
attempts should be made to eliminate them when getting final results.

Another more subtle issue is that a particular compilation / run may just
happen to have poor cache behaviour, depending on where code/stacks get
allocated ... Ho humm

Patrick.


------- End of Forwarded Message



Reply via email to