Aleksey, the patch looks good. Could you change "BLOCK_BASE_ALIGNMENT" to be "GC_BLOCK_BODY_ALIGNMENT"? BASE actually refers to the start of a block. BODY is the start of the allocatable area. (And same to BLOCK_BASE_ALIGN_MASK). Thanks.
I will commit after you make the modification. Thanks, xiaofeng On Fri, Apr 18, 2008 at 1:01 AM, Aleksey Shipilev (JIRA) <[EMAIL PROTECTED]> wrote: > > [ > https://issues.apache.org/jira/browse/HARMONY-5759?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12590102#action_12590102 > ] > > Aleksey Shipilev commented on HARMONY-5759: > ------------------------------------------- > > RC1 patch gives astonishing +30% boost to composite SciMark2 score. > > $ ../../builds/shade.r649000.CLEAN/bin/java -Xem:server > jnt.scimark2.commandline > > SciMark 2.0a > > Composite Score: 257.2925459369102 > FFT (1024): 324.4495281837533 > SOR (100x100): 364.95358549804746 > Monte Carlo : 61.79453602352396 > Sparse matmult (N=1000, nz=5000): 265.54295673436866 > LU (100x100): 269.7221232448578 > > java.vendor: Apache Software Foundation > java.version: 1.5.0 > os.arch: x86 > os.name: Windows XP > os.version: 5.1 > > 20:55:23 (189) /cygdrive/c/Harmony.new/test/Scimark > $ ../../builds/shade.r649000.GCB/bin/java -Xem:server > jnt.scimark2.commandline > > SciMark 2.0a > > Composite Score: 335.49067843718 > FFT (1024): 317.1346618181818 > SOR (100x100): 592.9092332307692 > Monte Carlo : 64.09633310436254 > Sparse matmult (N=1000, nz=5000): 270.69805822691814 > LU (100x100): 432.6151058056684 > > java.vendor: Apache Software Foundation > java.version: 1.5.0 > os.arch: x86 > os.name: Windows XP > os.version: 5.1 > > > > [drlvm][gc][performance] Alignment of GC blocks could improve the > performance of FP benchmarks > > > ---------------------------------------------------------------------------------------------- > > > > Key: HARMONY-5759 > > URL: https://issues.apache.org/jira/browse/HARMONY-5759 > > Project: Harmony > > Issue Type: Improvement > > Reporter: Aleksey Shipilev > > Attachments: vm-gc-blocksize-rc1.patch > > > > > > This issue defines the block alignment to be separated from object > alignment. This would allow to leave the object alignment at 4 and not to > waste precious memory, but on other hand will open the way for alignment of > GC blocks to 8 thus increasing the chances that data requiring 8/16 alignment > will layout in memory properly. > > -- > This message is automatically generated by JIRA. > - > You can reply to this email to add a comment to the issue online. > > -- http://xiao-feng.blogspot.com
