The way I was collecting heap stats had too much of an observer effect. Originally, it doubled the time of compiling a small program (4 GCs before taking the heap measurement). Even just doing 1 GC before taking the heap size increased wall clock time by 25%.
I think we need more careful thought on getting those metrics, so I removed them and replaced them with a TODO for now. Jason suggested we might use other jmx properties in the report. Maybe it would be best to just measured the tenured heap... http://gwt-code-reviews.appspot.com/1045801/show -- http://groups.google.com/group/Google-Web-Toolkit-Contributors
