From: Steve B. <Ste...@an...> - 2007-07-31 21:17:15
|
On 31/07/2007, at 11:04 PM, David P Grove wrote: > Steve, thanks for automating the search. My prime suspect is > 13187, but it would be nice to be sure. That experiment just finished. It's pretty clear that a number of the revisions regress specific benchmarks quite markedly. It would be great if, armed with this, we could now chase them down further. To help clarify things a little further, I'm now running the benchmarks all again [1], this time using 10th iteration replay advice (ie close to steady-state). The point of this is just that this is normally a lot less noisy since each benchmark execution will run exactly the same set of optimizations etc (the set was gathered by taking the best-performing 10th iteration run of 5 executions of the benchmark on 2.9.1). So that data should compliment the data I have already run [2]. Interestingly, initial data is quite different to [2], at least for antlr. Not sure how to explain that. :-/ Cheers, --Steve [1] http://cs.anu.edu.au/people/Steve.Blackburn/private/jikesrvm/ vole-2007-08-01-Wed-070146.time.txt [2] http://cs.anu.edu.au/people/Steve.Blackburn/private/jikesrvm/ vole-2007-07-31-Tue-214137.time.txt |