[CSC 335] Complex Math/Linear Algebra using Java

[Andrew Pounds]

So I took some time today to look back over the state of Java for 
numerical programming.  The "old reliables" of JAMA,  JLAPACK and COLT 
can still be downloaded (see the class webpage for references).  
Recognize -- some of these are OLD and were just abandoned because other 
languages VASTLY outperformed pure Java implementations.  Some of the 
web pages haven't been updated in 15 years.  One site showed a 3000x3000 
matrix multiply taking around 4 seconds in C/Fortran and 50 seconds in 
Java.  Java libraries built with Fortran/C back ends were closer to the 
pure C/Fortran values.

Newer implementations include the Apache Commons math classes as well as 
the JavaNumerics site from NIST.  It is clear that people REALLY want a 
good Java numerical toolkit as you will even find many webpages 
dedicated to it -- there is even a students attempt to wrap the Gnu 
Scientific Libraries with java code.  The rationale you will find for 
the continued development of these Java numerical libraries falls into 
two camps...

1. Java is the premier language for introducing CS (which was true 10 
years ago, but not so much now) and therefore we should abandon the 
older C and Fortran libraries for it because those languages are too 
limited.

2. It may be slow, but eventually computers will be so fast nobody will 
know the difference.

I contend that Java is a great language -- but it has its place. In my 
opinion Java wasn't designed for HPC.  As far as the second point goes, 
machines will get faster, but by definition the other programming 
languages will too.  I have heard the same rationale used to defend 
Python -- which I love -- but in which I would never write a "production 
level" HPC system that had to be really fast.

I will be curious to see how Julia (a newer HPC language) does in this 
mix as it is built from C/Fortran and designed to take advantage of the 
aspects of those languages that make them ideal for HPC applications.


For those that are curious Mathematica and Matlab were written in C/C++ 
and Java (you can guess which parts are in C).  Octave (which is 
available on cobra and is similar to Matlab) was written in C/C++ and 
Fortran.


With all that said, if anyone needs an example of how to do complex math 
and linear algebra with Java let me know and I'll see what I can cobble 
together.



-- 
Andrew J. Pounds, Ph.D.
Professor of Chemistry and Computer Science
Mercer University,  Macon, GA 31207   (478) 301-5627
http://faculty.mercer.edu/pounds_aj