Weird profiling result of mwait_idle() function by VTune.
I used VTune 3.0 to profile the Linuxnetwork stack activities(RHEL 4.0 with 2.6 kernel). The NIC driveri used is 6.0.60 forIntelgigabit dual-port NIC.
I found the total clock ticks usedby the mwait_idle() function is unbelieveably as high as 50% of total clock ticks when I enabled hyper-threading for a dual Xeon processor platforms and the CPI for this mwait_idle() is about 4000. I did the same testing on a single Xeon processor with hyper threading enabled. The mwait_idle() only used 5% of total clock ticks and the CPI is about 1000.
I understand why CPI of mwait is so high (run as fewer instructions as possible).
But mwait_idle() should not be take half of clock ticks? If so, this means there is alwayshalf ofCPUsbeing idle for this dual-cpu platform? Right?
Anyonehas any clueabout this?
Message Edited by email@example.com on 08-17-200505:10 PM