Vortrag Susan Flynn-Hummel, 14.8.98, 14.00 Uhr

Maria Cherry Maria Cherry <maria@par.univie.ac.at>
Mon, 3 Aug 1998 09:37:35 +0200 (MET DST)


                          UNIVERSITAET WIEN 
          INSTITUT FUER SOFTWARETECHNIK UND PARALLELE SYSTEME
                            gemeinsam mit 
                                VCPC 
           EUROPEAN CENTRE FOR PARALLEL COMPUTING AT VIENNA 


      EINLADUNG ZU EINEM VORTRAG IM RAHMEN DES INSTITUTS-KOLLOQUIUMS:
                
         
                                Jalapeno
                   
                          Dr. Susan Flynn-Hummel
			    Java Tools Group*
 		   IBM T. J. Watson Research Center
      
 		               
               ZEIT: Freitag, 14. 8. 1998, 14.00 Uhr c.t.
                                             
                                             
       ORT: Institut fuer Softwaretechnik und Parallele Systeme
                  1090 Wien, Liechtensteinstrasse 22, 
                         Seminarraum, Mezzanin


Abstract


The goal of the Jalapeno project at IBM T. J. Watson Research Center is
to push the envelope of Java Virtual Machine (JVM) technology.  In
particular, to develop a flexible infrastructure for experimenting with
and building scalable high-performance Java servers to be run on
Shared-memory Multi-Processors (SMPs).  Our design thrust is to get
something working as early as possible then to optimize pieces that are
observed to be performance bottlenecks.  Our initial target platform is
the IBM PowerPC.

The Jalapeno JVM is "purer-that-pure" Java.  All internal
data-structures are Java objects.  Native methods in Sun's JDK are
reimplemented in Java ("no native code" TM) as they are needed.  This
design choice brings with it several advantages as well as "research
opportunities".  Obvious benefits are those of using a modern
object-oriented language for developing large-scale system software,
and insights into the generation of high-performance code for such a
system.  Novel use of the reflection and introspection capabilities of
Java allowed early construction of debuggers and a JVM "boot image"
builder (all written in Java).  And, although we are unwilling to
sacrifice performance for portability, by isolating system dependent
functionality, the choice of using Java should facilitate porting to
other architectures.

Jalapeno does not interpret bytecodes, instead "fast" compilers
translate them directly to machine code.  Here, the term "fast" is used
in three technical senses: 1) fast development, 2) fast code generation
and 3) fast code execution, leading to three separate research
efforts.  The first, a base-line compiler, provided us with a basis for
understanding the interaction of user Java programs and of the JVM
itself, for experimenting with differing algorithms for garbage
collection, multi-threading etc, and a tool for debugging (both user
code and our other compilers).  This compiler mimics Sun's stack-based
Java VM specification.  It is slow and the compiled code barely keeps
pace with the JDK interpreter.  A second effort, which will take
greater advantage of the PowerPC's registers, is underway to improve
both generation and execution times.  The third compiler effort aims at
dynamically producing highly optimized code for methods (or blocks)
that are observed at runtime to be bottlenecks.

*Participants of the Java Tools Jalapeno project include:  Bowen
Alpern, Dick Attanasio, John Barton, Philippe Charles, Jong-Deok Choi,
Tony Cocchi, Susan Flynn-Hummel, Chris Laffra, Derek Lieber, Mark
Mergen, Ton Anh Ngo, Vivek Sarkar, Janice Shepherd, David Shields, 
Steve Smith, Harini Srinivasan, Micheal Travers, and John Whaley.