> On 29 Jan 2016, at 17:43, Hamlin Li <huaming...@oracle.com> wrote: > > > > On 2016/1/29 20:53, Paul Sandoz wrote: >>> On 29 Jan 2016, at 13:43, Hamlin Li <huaming...@oracle.com> wrote: >>> >>> Hi Paul, >>> >>> Sorry for delayed response, have been occupied by other higher priority >>> task. >>> Thanks for your review, I agree with you that your second approach is >>> better. >>> New webrev: http://cr.openjdk.java.net/~mli/8076458/webrev.01/ >>> >> The changes to the data providers look ok. >> >> Would you mind splitting out the tests between StreamTestData<Integer> and >> StreamTestData<Integer>.small as outlined in 2) below. That way for the >> non-eplosive stuff we can still crunch on larger data without much of a slow >> down. > Hi Pual, > > Yes, you're right, it does not slow down too much, it cost 15.553 seconds > after the first revision(webrev.01), and it cost 16.064 after the second > revision(webrev.02). > Please check the webrev: http://cr.openjdk.java.net/~mli/8076458/webrev.02/ > <http://cr.openjdk.java.net/~mli/8076458/webrev.02/>
+1, reviewed. Do you need me to push it? >>> Below are times cost for different ops: >>> total :169.996 >>> testOps only :108.988 >>> testIntOps only :23.865 >>> testLongOps only :22.326 >>> testDoubleOps only :16.944 >>> so, I build small data providers for each of them. >>> >> Ok, and i suspect/hope it drops by at least an order of magnitude with your >> changes applied. > Yes, it cost 15.553 seconds after the first revision(webrev.01). Great. >> Out of curiosity i wonder what the times would be if using parallel GC >> rather than G1. > With different GC options after second revision(webrev.02): > -UseParallelGC: elapsed time (seconds): 16.047 > +UseParallelGC: elapsed time (seconds): 13.263 > -UseG1GC: elapsed time (seconds): 16.612 > +UseG1GC: elapsed time (seconds): 16.998 > -UseParallelOldGC: elapsed time (seconds): 16.039 > +UseParallelOldGC: elapsed time (seconds): 14.297 > Ok, so i suspect in the case of when your patch is not applied the difference is greater in absolute terms and G1 in a sense might be causing such memory intensive tests to slow down. Paul.
