The NINJA Project: Making Java Work for High Performance Numerical Computing

When Java(TM) was first introduced, there was a perception (properly founded at the time) that its many benefits came at a significant performance cost. In few areas were the performance deficiencies of Java so blatant as in numerical computing. Our own measurements, with second-generation Java virtual machines, showed differences in performance of up to one hundred-fold relative to C or Fortran. The initial experiences with such poor performance caused many developers of high performance numerical applications to reject Java out-of-hand as a platform for their applications. Despite the rapid progress that has been made in the past few years, the performance of commercially available Java platforms is not yet on par with state-of-the-art Fortran and C compilers. Programs using complex arithmetic exhibit particularly bad performance. Furthermore, current Java platforms are incapable of automatically applying important optimizations for numerical code, such as loop transformations and automatic parallelization. Nevertheless, our thesis is that there are no technical barriers to high performance computing in Java. To prove this thesis, we have developed a prototype Java environment, called Numerically INtensive JAva (NINJA), which has demonstrated Fortran-like performance can be obtained by Java on a variety of problems. We have successfully addressed issues such as dense and irregular matrix computations, calculations with complex numbers, automatic loop transformations, and automatic parallelization. Moreover, our techniques are straightforward to implement, and allow reuse of existing optimization components already deployed by software vendors for other languages, lowering the economic barriers to Java's acceptance. The next challenge for numerically intensive computing in Java is convincing developers and managers in this domain that Java's benefits can be obtained 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 excellent numerical performance, our techniques will allow vendors and researchers to quickly deliver high performance Java platforms to program developers.