Performance Counters Library - design concepts

The rabbit design concepts

Intel Pentium and Pentium Pro/II/III families of processors

time stamp counter (64-bit cycle counter)
two performance-monitoring counters (40-bit event counter)

AMD Athlon family of processors

time stamp counter (64-bit cycle counter)
four performance-monitoring counters (48-bit event counter)

select two/four events to monitor, processor will increment counters as events occur
periodic sampling of the counters, driven by an interval timer
time-shared rotation through a list of event pairs, changing the selection at the sampling points

pair = 2 or 4 events, as appropriate

use predefined groups of event pairs, or select your own pairs
summary output, or time-stamped data for plotting
no modifications to the target program are required
especially suited for non-interactive environments
lots of examples

The PMC library design concepts

Decisions about "when to measure" are placed directly in the program.
Decisions about "what to measure" are delayed until runtime.
Measurements are taken as cleanly as possible.
As many options are granted as the processor and operating system allow.
The data structures and interfaces are machine-independent as far as possible.
Measurement data is available to the program.

The current implementation

Linux 2.[024].x kernels
/dev/pmc loadable module
no other kernel modifications

measurements are attached to the processor, not the process
for best results, take measurements on a single-processor system with no other users