perfctr For Linux

Limitations in the unregistered version

• Kernels older than 2.4.16 are not supported since perfctr-2.6. You can use the previous stable series, perfctr-2.4, if you must use an older kernel, but this has several limitations:
• Older kernels do not support AMD64 (x86-64).
• The performance counters in hyper-threaded P4s/Xeons cannot be used with kernels older than 2.4.15. You'd have to disable hyper-threading or SMP, or restrict yourself to TSC sampling.
• No profiling using counter overflow interrupts, except in 2.4.10 and newer kernels, and some early 2.4-ac/redhat kernels.
• Application code compiled for perfctr-2.4 is not compatible with perfctr-2.6, and vice versa.
• The perfctr-2.4 series does not support 2.6 kernels. Some of these limitations may be fixable. Contact the author if you are willing to fund development in this direction.
• The performance counter interrupt facility requires SMP or uniprocessor APIC support. In the latter case, the BIOS must be reasonably non-buggy. Unfortunately, this is often not the case.
• Neither the kernel driver nor the sample user-space library attempt to hide any processor-specific details from the user.
• This package makes it possible to compute aggregate event and cycle counts for sections of code. Since many x86-type processors use out-of-order execution, it is impossible to attribute exact event or cycle counts to individual instructions.
• Centaur WinChip C6/2/3 support requires that the TSC is disabled.