This is question that involves the Intel Performance
Primitives Library (IPPL), but it is more a question about
timing on a Pentium 4 with hyperthreading using the IPPL.
I have an application that uses the IPPL on a Windows XP
system running a Pentium 4 with hyperthreading turned on,
i.e it has two logical processors, and one physical processor.
When we time calculations using the Windows GetProcessTimes
interface we get the following curious result.
When we do a calculation that shows 100% total CPU usage
(i.e. 100% on both CPUs) we get a time that is very close to
twice wallclock time. That is expected.
When we do a different calculation that shows 60% total CPU usage
(according to the Windows performance monitor) we get agreement
with wallclock time to say three figures. Why is this happening?
I would have expected that the total process time would differ
from wallclock by say 20%. In other words, I would have expected
a gradual increase in process time as the CPU utilization
goes from 50% to 100%. But at 60% my data shows no change relative
to wallclock time.
I would appreciate any insight into this matter.
