the ninja project making java work for high performance computing

TheNINJA Project: Making JavaWork
for High PerformanceNumerical Computing
Jos? eE. Moreira Samuel P. Midkiff Manish Gupta
IBM ThomasJ. Watson Research Center
Pedro Artigas
School of Computer Science, CarnegieMellon University
Peng Wu GeorgeAlmasi
Department of Computer Science, University of Illinois
1 Introduction
When Java
(TM)
was ?rst introduced, there was a perception (properly founded at the time) that its many
bene?ts came at a signi?cant performance cost. In few areas were the performance de?ciencies of Java so
blatant as in numerical computing. Our own measurements, with second-generation Java virtual machines,
showeddifferencesinperformanceof uptoonehundred-fold relativetoCor Fortran. Theinitial experiences
with such poor performance caused many developers of high performance numerical applications to reject
Javaout-of-hand asaplatformfor their applications.
Despite the rapid progress that has been made in the past few years, the performance of commercially
availableJavaplatformsisnot yet onpar withstate-of-the-art FortranandCcompilers. Programsusingcom-
plex arithmetic exhibit particularly bad performance. Furthermore, current Java platforms are incapable of
automatically applying important optimizations for numerical code, such asloop transformations and auto-
matic parallelization [12]. Nevertheless, our thesisisthat thereareno technical barriersto high performance
computing in Java. To prove this thesis, we have developed a prototype Java environment, called Numer-
ically INtensive JAva (NINJA), which has demonstrated Fortran-like performance can be obtained by Java
on avariety of problems. Wehavesuccessfully addressed issues such asdenseand irregular matrix compu-
tations, calculations with complex numbers, automatic loop transformations, and automatic parallelization.
Moreover, our techniques are straightforward to implement, and allow reuse of existing optimization com-
ponents already deployed by software vendors for other languages [9], lowering the economic barriers to
Java’s acceptance.
Thenext challenge for numerically intensive computing in Javaisconvincing developers and managers
in this domain that Java’s bene?ts can beobtained with performance comparable with highly tuned Fortran
and C. This treatise contributes to overcoming that challenge. Once it is accepted that Java performance
is only an artifact of particular implementations of Java, and that there are no technical barriers to Java
achieving ex cellent numerical performance, our techniques will allow vendors and researchers to quickly
deliver high performance Javaplatforms to programdevelopers.
Thecorresponding author may bereached at jmoreira@http://wendang.chazidian.com, or P.O. Box 218, Yorktown Heights, NY 10598.
1