Running WL 6.1 SP2 on Sun's new JVM 1.4

Question

Hi all,
First a bit of background to my request.
We have an WL application running on Solaris that simply must have a very large
EJB cache since the time taken to load each EJB and its dependent objects is very
large. We frequently do operations over large sets of the objects.
Because of this, these objects end up being tenured to the old part of the JVMs
heap, I'm talking up to a gigabyte+ of data. We have tried every option to the
garbage collector on Sun's 1.3.1 JVM to improve performance, but cannot get rid
of the fact that a major garbage collection will take place once every minute
or so, and when it does, it takes about 2.5 secs for each 40 meg in the old heap,
in short a horrible pause and all processors are locked out for that duration.
It's a highly interactive application. We have tried using -Xincgc for incremental
garbage collection but major collections still take place, although less often,
still enough to pose a serious problem, increasingly so as the heap gets larger.
This is a real blow for enterprise Java computing with large caches, the JVMs
to date are just not designed for large multi-threaded server side apps
JVM 1.4 is said to have addressed some of this with parallel garbage collection,
which would mean our WL application threads could keep running while GC is taking
place. So we tried booting up WL 6.1 on JDK 1.4. We did install the 64 bit version
since we are running 64 bit Solaris.
With the 32 bit option to the JVM, we get a static initialiser problem the second
we start it, with the 64 bit option, we get cannot find execv() or some such function.
So my question is, has anybody successfully run their WL 6.1 on the new JDK 1.4,
or any version of WL for that matter. If so, did you have to do anything specific,
patches, OS reconfigs etc to get it to run? If anybody from BEA is reading, is
BEA likely to certify WL 6.1 on JVM 1.4?
No doubt somebody will probably mention giving JRockit a go, a server side JVM
designed from the ground up for the type of app I'm talking about. BEA used it
recently to get their latest benchmark with WL 7.0 beta.
We did download this and try it. Unfortunately, JRockit on Solaris is only at
a 3.0 release compared to the later 3.1 release on NT and Linux. Firstly the 3.0
starts up way slower than the 3.1 release. But that's ok, one can live with a
slower startup. Once started up on Solaris (2.8), it does run quickly alright,
but after a while it just simply hangs, with no errors or anything, the process
stays alive with 0% CPU, and does nothing, you can't connect to WL or anything,
you just have to kill it. Other groups with my organization experienced the same
thing on Solaris. We wrote to JRockit who said yes they have similar problems
with Solaris in their own labs but could not say when there will be another release.
We simply have to deploy on Solaris, so JVM 1.4 right now seems my only hope.
Any info greatly received.
Regs,
Paul

Answer 1

It should be fairly easy for you to track down who's calling System.gc().
I believe the Sun JDK has the java.lang.* sources bundled with the jvm (or perhaps it's
a separate download.) Add a Thread.dumpStack() to System.java, re-compile, and prepend
your new version to the bootclasspath. (type java -X for more info on setting the
bootclasspath)
-- Rob
Paul Hammond wrote:
Thanks a lot Danny, will check that.
Actually I should mention that using the non standard
-XX:+DisableExplicitGC option to the JVM got rid of practically all the full GCs,
seems like probably WebLogic or possibly some other third party library periodically
calls System.gc() and this really kills it if you have a large old heap.
Regs,
Paul
Danny Foncke <[email protected]> wrote:
There's something at :
http://www.genuitec.com/products.htm#weblogic_jdk14
Danny
Paul Hammond wrote:
Hi all,
First a bit of background to my request.
We have an WL application running on Solaris that simply must havea very large
EJB cache since the time taken to load each EJB and its dependent objectsis very
large. We frequently do operations over large sets of the objects.
Because of this, these objects end up being tenured to the old partof the JVMs
heap, I'm talking up to a gigabyte+ of data. We have tried every optionto the
garbage collector on Sun's 1.3.1 JVM to improve performance, but cannotget rid
of the fact that a major garbage collection will take place once everyminute
or so, and when it does, it takes about 2.5 secs for each 40 meg inthe old heap,
in short a horrible pause and all processors are locked out for thatduration.
It's a highly interactive application. We have tried using -Xincgcfor incremental
garbage collection but major collections still take place, althoughless often,
still enough to pose a serious problem, increasingly so as the heapgets larger.
This is a real blow for enterprise Java computing with large caches,the JVMs
to date are just not designed for large multi-threaded server sideapps
JVM 1.4 is said to have addressed some of this with parallel garbagecollection,
which would mean our WL application threads could keep running whileGC is taking
place. So we tried booting up WL 6.1 on JDK 1.4. We did install the64 bit version
since we are running 64 bit Solaris.
With the 32 bit option to the JVM, we get a static initialiser problemthe second
we start it, with the 64 bit option, we get cannot find execv() orsome such function.
So my question is, has anybody successfully run their WL 6.1 on thenew JDK 1.4,
or any version of WL for that matter. If so, did you have to do anythingspecific,
patches, OS reconfigs etc to get it to run? If anybody from BEA isreading, is
BEA likely to certify WL 6.1 on JVM 1.4?
No doubt somebody will probably mention giving JRockit a go, a serverside JVM
designed from the ground up for the type of app I'm talking about.BEA used it
recently to get their latest benchmark with WL 7.0 beta.
We did download this and try it. Unfortunately, JRockit on Solarisis only at
a 3.0 release compared to the later 3.1 release on NT and Linux. Firstlythe 3.0
starts up way slower than the 3.1 release. But that's ok, one can livewith a
slower startup. Once started up on Solaris (2.8), it does run quicklyalright,
but after a while it just simply hangs, with no errors or anything,the process
stays alive with 0% CPU, and does nothing, you can't connect to WLor anything,
you just have to kill it. Other groups with my organization experiencedthe same
thing on Solaris. We wrote to JRockit who said yes they have similarproblems
with Solaris in their own labs but could not say when there will beanother release.
We simply have to deploy on Solaris, so JVM 1.4 right now seems myonly hope.
Any info greatly received.
Regs,
Paul