Nearly 10 years ago, we found an interesting way of characterizing program behavior (particularly CPU/MEM bounds) by breaking code semantics, substituting array accesses with scalar accesses and observing reaction. There we could predict CPU/MEM bounds 3 orders of magnitude faster than simulators and much more precise than hardware counters ^{Fur2004 FOTP2004}. Later, we used reactions to optimizations to find similar programs when predicting optimizations ^FT2010.

We are gradually adding these techniques to cM and Alchemist, and plan to use it to analyze parallel applications for bottlenecks (removing or adding threads), etc - collaborations and help are very welcome!