JDK-4906307 : Performance of launching demos with -server -Xcomp is much slower in tiger
  • Type: Bug
  • Component: client-libs
  • Sub-Component: 2d
  • Affected Version: 5.0
  • Priority: P4
  • Status: Closed
  • Resolution: Not an Issue
  • OS: solaris_10
  • CPU: sparc
  • Submitted: 2003-08-13
  • Updated: 2003-10-15
  • Resolved: 2003-10-13
Related Reports
Relates :  
Relates :  
Relates :  
Description
When launching Java2D and SwingSet2 demos in j2sdk1.5.0 b15 in Solaris 10 (B29 and B34) machine with -server -Xcomp flags, I noticed that the performance is much slower compared with that in j2sdk1.4.2_01 B04.

STEP TO REPRODUCE
1. Install j2sdk1.5.0 b15 & j2sdk1.4.2_01 b04 in Sol10
2. Launch the Java2D demo with -server -Xcomp flags and time the completed launching time with 1.5.0 b15.
3. Launch the Java2D demo with -server -Xcomp flags with 1.4.2_01 b04 and time the completed launching time.
4. Following the same procedure for SwingSet2 demo too.
5. The results of launching Java2Demo and SwingSet2 demos with -server -Xcomp flags is much slower in 1.5.0 b15 compared with 1.4.2_01 b04 in Sol10 machine.

###@###.### 2003-08-13

Comments
EVALUATION ###@###.### 2003-09-11 According to jvmstat data for Java2D the performance of server JVM 1.5.0 is improved when used with JDK 1.4.2: 1. It has less failed compilations. 2. It spends the same (or less) time to compile more bytecodes. The reason of slowdown in JDK 1.5.0 is increased (>10%) numbers of methods been compiled: jdk1.4.2_jvm1.4.2 jdk1.4.2_jvm1.5.0 jdk1.5.0_jvm1.5.0 hotspot.ci.native.compiles 349 349 363 hotspot.ci.osr.compiles 1 1 2 hotspot.ci.standard.bytes 575222 585729 662065 hotspot.ci.total.bailout 146 34 26 hotspot.ci.total.compiles 8023 8226 9099 hotspot.ci.total.invalid 35 0 0 hotspot.ci.total.time 112627140767 111604414876 129604662405 hotspot.rt.cl.bytes.loaded 2109056 2170928 2331992 hotspot.rt.cl.classes.loaded 1743 1743 1848 hotspot.rt.hrt.ticks 264289262070 272758923408 285865197794 hotspot.rt.application.time 262274979354 272407660188 I attached outputs of -XX:+PrintCompilatiop for latest 1.4.2_02 and 1.5.0. (I used the same Java2Demo.jar from 1.5.0 b10). % wc comp.1.4.2.noinl.sort 8914 8914 378761 comp.1.4.2.noinl.sort % wc comp.1.5.0.noinl.sort 10046 10046 420731 comp.1.5.0.noinl.sort 1. sun.awt.font was replaced with sun.font: % fgrep sun.awt.font comp.1.4.2.noinl.sort|wc 191 191 8754 % fgrep sun.font comp.1.5.0.noinl.sort|wc 511 511 19730 2. New code added in java.nio % fgrep java.nio comp.1.4.2.noinl.sort|wc 67 67 2544 % fgrep java.nio comp.1.5.0.noinl.sort|wc 246 246 9351 3. Added calls to java.util.regex % fgrep java.util.regex comp.1.4.2.noinl.sort|wc 0 0 0 % fgrep java.util.regex comp.1.5.0.noinl.sort|wc 80 80 3153 4. Added calls to java.util.logging % fgrep java.util.logging comp.1.4.2.noinl.sort|wc 0 0 0 % fgrep java.util.logging comp.1.5.0.noinl.sort|wc 81 81 3411 And so on. I am passing this bug to java/classes_fontprop since java.font is a "winner". more acutely speaking, the owner of the "winer" is "2d" ###@###.### 2003-10-09 =============================== The synopsis of this bug could be rewritten as "more classes are loaded in 1.5" since -server -Xcomp start-up time is obviously driven by that. If this were the normal way client apps start-up we'd rewrite everything in C, but its not, and in fact the apps start-up faster because of the new Java code. At least that's the case for the sun.font classes. I think some of the classloading graphs of the performance team already highlighted the majority of these issues. I believe the loading of logging classes (by AWT???) is being fixed. I have no idea about the regex classes. Also it could be that the server compiler in 1.5 is slower than that in 1.4.2, this may be contributing in some part. ###@###.### 2003-10-13 ============================
13-10-2003