When I'm using oprofile on AMD Phenom(tm) 9750 Quad-Core Processor machine, and the linux version is 2.6.24. The event lists used are L2_CACHE_MISS and REQUESTS_TO_L2, when I change the count of the event property, an interesting thing happens:
When the count values are both set to 500, then it shows 10%~20% CPU utilization ratio measured by 'top' command.
When the count values are both set to 3000, then it shows 50% or so CPU utilization ratio.
When the count valuse are changed to 6000, the ratio reaches 100%.
Theoretically, the greater the count value is, the less overhead oprofiled brings, but what's the wrong?