On Wednesday 21 May 2008 11:47:34 Adrian Brock wrote: I've attached the profile for beta4 and rev73546 to the task. Hopefully the information you requested is located in the caller/ directory. Otherwise just let me know and I'll change the report. Best regards, Jesper

On Wed, 2008-05-21 at 11:47 +0200, Adrian Brock wrote: I don't fully understand the file that Stale added to JBVFS-26 It does show an increase in the number of classloader attempts to invoke getChild() since jboss-beta4. Overall this looks to be spending 5 times the amount of time per request inside the VFS layer. e.g. I see VFSClassLoaderPolicy.getProtectionDomain() as beta4 = 78,497 calls - 235 ms) now = 127,067 calls - 2,494 ms I get the number 5 by taking the 10 fold increase in time and dividing by the nearly two fold increase in number of invocations. What I can't understand is why getProtectionDomain() has this many invocations. The AS bootstrap has only 13,000 classes loaded (including JDK classes). I guess I'm reading this wrong. If that numnber really was true 90% of these invocations would lead to exceptions saying duplicate invocation of ClassLoader.defineClass(). Since you can't define a class twice and getProtectionDomain() is only invoked from there. ;-) One thing I did notice, is the duplicate invocation of findChild() during the defineClass() process. It's sort of reasonable because it is locating different parts of the VFS structure (e.g. the class file versus the root for the code source) but its also a bit stupid since this information can be cached the first time you do it. Although it doesn't make much difference in my subjective testing < 2 seconds (I'm not even sure that number is real since I only ran it once), I committed it anyway. Maybe somebody else will see some benefit? See http://jira.jboss.com/jira/browse/JBCL-12 which will be in the next classloader release. -- xxxxxxxxxxxxxxxxxxxxxxxxxxxx Adrian Brock Chief Scientist JBoss, a division of Red Hat xxxxxxxxxxxxxxxxxxxxxxxxxxxx