Status: Accepted
Owner: erik.corry
Labels: Type-Bug Priority-Medium
New issue 1934 by erik.corry: Permute benchmark is very flaky
http://code.google.com/p/v8/issues/detail?id=1934
Starting somewhere around revision 9415 the Permute benchmark (part of
BenchPress) shows very variable results. Starting around there around 1 in
3 test dashboard runs show a score of slower than 180. Before that there
had been no runs that were that slow since 6216.
Prime suspect:
9415: Add dynamic stack frame alignment to optimized functions with
untagged doubles on the stack.
Around 10202 the average performance got worse. There is no obvious
suspect for this.
Recently the average performance has improved, but the variation between
runs has persisted. The prime suspect for the improvement is
10300: ...I also refactored and improved the code for smi-only arrays.
There have been no scores worse than 200 since 10295.
The hash randomization changes starting at 10330 seem to have had no effect
on this issue.
--
v8-dev mailing list
[email protected]
http://groups.google.com/group/v8-dev
