Abstract

Java is often accused of being too slow for serious programming, especially for scientific problem solving. However, we found that for a large-scale geophysical application, Java code compiled with current just-intime compilers runs slower than Fortran by a factor of at most four, on both a shared-memory parallel machine (SGI Origin000) and a distributed-memory parallel machine (IBM $P/). The moderate slowdown is easily offset by the following advantages: (a) object-oriented Java code is easier to maintain and reuse than Fortran code, (b) Java code is fully portable, even among parallel computers with different memory models. Furthermore, better compiler technology is on the horizon, which will narrow the performance gap even more.

Citations