Re: JIT and class loaders

Blair and Richard,

Thanks for the information. What I have noticed is that
the first touch on a class takes a while. It seems to
matter much less when we use the system class loader.
However, the benefits we get from a user class loader
are substantial. If necessary we can set up a task to
go through and pre-open pages, but that is a pain for
our developers. We did not need to do this with Tomcat
3.2.3, but it used the system class loader much more
frequently than 4.0.1.

>From what you say the JIT should not have as much
impact as what I am seeing. I don't have any hard and
fast numbers, but about 10 seconds first touch, 1 on
subsequent calls is typical. Your comment prompted
me to double check to make sure that Jasper (the
internal Tomcat compiler) was not being invoked,
except on a change. It looks OK.

I based the -O question on the following comment:

Use the -O javac option

The javac compiler may in-line methods if the -O option is selected. In
order for a method to be in-lined it must be a final method (static,
final, or protected). Final methods can not be subclassed, which is the
key to allowing in-lining. Method in-lining is not supported in the
AS/400 JVM other than through the -O option.

I found that in this document:

http://www.iseries.ibm.com/developer/java/faq/perffaq.html

David Morris


>>> cujo@us.ibm.com 01/08/02 02:26PM >>>

FWIW:  About a year ago, the -O option was a total no-op.  Something
might
have been added since then, but its unlikely.  And that information
applies
to the Sun supplied javac command.  I assume platforms could do
various
things but don't know the intenal logistics of exactly what would be
done
or how personally.

Richard D. Dettinger
AS/400 Java Data Access Team

Democracy's enemies have always underestimated the courage of the
American
people.
It was true at Concord Bridge.  It was true at Pearl Harbor.  And it
was
true today.

         Rochester Post-Bulletin
         Tuesday September 11, 2001


Blair Wyman/Rochester/IBM@IBMUS@midrange.com on 01/08/2002 03:11:30 PM

Please respond to java400-l@midrange.com

Sent by:    java400-l-admin@midrange.com


To:    java400-l@midrange.com
cc:
Subject:    Re: JIT and class loaders




David Morris wrote:

>Group,
>
>I searched the iSeries FAQs and can't find a clear description of
when
>the JVM supports Jar and Class files created using CRTJVAPGM. I am
>assuming that if I use a custom class loader (almost any Servlet
>container) that the iSeries JVM will have to fall back to the JIT
>optimizations if available. Is that true? Also, is inlining possible
>with the JIT compiler and does the -O compile option have any affect?

Currently, any class loaded via a so-called "user classloader" will
be processed (by default) using the JIT.  This is an artifact of the
way our JVM finds the pre-compiled result of calling CRTJVAPGM, which
is using the file's inode.  When a user classloader is involved, the
JVM never gets to see the actual file (either by name or by inode),
and so there's no easy way to get to the JVAPGM.

The terminology "fall back" is a little misleading -- the JIT
performance
is fast approaching the performance of CRTJVAPGM'd code, right on up
to
level 40.  In some cases, even the most aggressive "static" analyses
and
optimizations, done by CRTJVAPGM at high optimization levels, aren't
as
effective as the optimizations possible using the JIT's complete
knowledge of the runtime environment.  (Static optimizations, due to
the dynamic nature of Java, have to be "guarded" -- the JIT is under
no such restriction.)

So, yes, the JIT can and does perform inlining, and in fact can do so
when CRTJVAPGM cannot.

The -O compile option might have some effect, but that's obviously all
happening before either the JIT or CRTJVAPGM get any chance, since
it's
an option on the javac command.  (javac takes source code to
bytecode.)
I haven't played with it much, but together with javap -c (to
disassemble
a classfile) it could be interesting to see what it "tries."  :)

HTH.

-blair

Blair Wyman -- iSeries JVM -- (507) 253-2891
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
"I was born not knowing, and have had only a
little time to change that here and there."   -- Richard P. Feynman




                      "David Morris"
                      <David.Morris@plu        To:
                      <JAVA400-L@midrange.com>
                      mcreek.com>              cc:
                      Sent by:                 Subject:  JIT and class
                      loaders
                      java400-l-admin@m
                      idrange.com


                      01/08/2002 01:17
                      PM
                      Please respond to
                      java400-l