> These experiments should be run twice: one set with an event that is > only triggered in the user-space (e.g. FP op counting) and a second > set an event that occurs both in kernel and user space.
Some architectures do some floating point operations in the kernel (e.g. ia64 ... gcc generates code that uses the FPU for integer multiply). So you may not be able to use a strictly "kernel count must be zero" for FP operation events. Perhaps this can be inverted ... find an event that is only triggered by a privileged instruction ... so you can see some "kernel only" counting happening. Still have the problem of determining whether the value you see is sane. You could compare results across different time intervals (while running the same load) and check that the counts scale correctly. -Tony _______________________________________________ perfmon mailing list [email protected] http://www.hpl.hp.com/hosted/linux/mail-archives/perfmon/
