Generalized Value Profiling and Instruction Profiling in DCPI
The Digitial (now Compaq) Continuous Profiling Infrastructure is sampling-based profiler that, without any recompilation, profiles whole systems with low overhead.
This talk will describe recent extensions to DCPI to support generalized value profiling and experience using profile-me performance counters in the Alpha 21264a.
Generalized value profiling periodically interprets a run of instructions, gathering frequently computed "values" for each instruction. One can gather both tradition values, e.g., the sum computed by an add instruction, or other values, such as the addresses of other instructions that might cause an instruction to incur a replay trap on the 21264.
The talk will motivate the need for generalized value profiling and profile-me performance counters with examples where DCPI was used to speedup the fastest Java Virtual Machine (see http://www.spec.org/osg/jvm98) and the inner loop of AltaVista's search engine.
Mark Vandevoorde is one of many co-authors of DCPI. Formerly at Compaq's System Research Center, Mark is now at AltaVista. He has a PhD from M.I.T.