Affiliations: IBM Thomas J. Watson Research Center, Yorktown
Heights, NY 10598-0218, USA. Tel.: +1 914 945 3018; Fax: +1 914 945 4270;
E-mail: {jmoreira,smidkiff,mgupta,pengwu,gheorghe}@us.ibm.com | School of Computer Science, Carnegie Mellon
University, Pittsburgh, PA 15213-3891, USA. E-mail: artigas@cs.cmu.edu
Abstract: When Java was first introduced, there was a perception that its many
benefits came at a significant performance cost. In the particularly
performance-sensitive field of numerical computing, initial measurements
indicated a hundred-fold performance disadvantage between Java and more
established languages such as Fortran and C. Although much progress has been
made, and Java now can be competitive with C/C++ in many important situations,
significant performance challenges remain. Existing Java virtual machines are
not yet capable of performing the advanced loop transformations and automatic
parallelization that are now common in state-of-the-art Fortran compilers. Java
also has difficulties in implementing complex arithmetic efficiently. These
performance deficiencies can be attacked with a combination of class libraries
({\it packages}, in Java) that implement truly multidimensional arrays and
complex numbers, and new compiler techniques that exploit the properties of
these class libraries to enable other, more conventional, optimizations. Two
compiler techniques, {\it versioning} and {\it semantic expansion}, can be
leveraged to allow fully automatic optimization and parallelization of Java
code. Our measurements with the NINJA prototype Java environment show that Java
can be competitive in performance with highly optimized and tuned Fortran
code.
