As of RedHat 6.4, the default setting for the kernel is to set the CR4.PCE bit, which enables user-mode code to execute the RDPMC instruction. This allows access to the core hardware performance counters at about 1% of the overhead of going through a kernel driver.
Unfortunately VTune seems to think that this is a bad idea, since it clears CR4.PCE on exit. This is a serious problem, as it requires executing kernel code to fix.
Is there any way to prevent VTune from clearing CR4.PCE on exit? Ideally it should read the value of CR4.PCE on startup and restore that value on exit...
Intel(R) VTune(TM) Amplifier XE 2013 Update 17 (build 353306) Command Line Tool
Copyright (C) 2009-2014 Intel Corporation. All rights reserved.