SciMark 2.0 Java Numerical Benchmark

    Roldan Pozo, Bruce Miller

    NIST

SciMark 2.0 is a composite Java benchmark measuring the performance of numerical kernels occurring in scientific and engineering applications.
It consists of five kernels which typify computational routines commonly found in numeric codes: Fast Fourier Transforms (FFTs), Jacobi Successive Over-relaxation (SOR), Sparse matrix-multiply, Monte Carlo integration, and dense LU matrix factorization.

(See http://www.math.nist.gov/scimark for further information and latest updates.)

INSTALLATION

Unpack the contents of archive into a subdirectory on your CLASSPATH. Be sure to keep the directory structure of the file contents.

COMPILING THE BENCHMARKS (optional)

From the directory above this one, issue the command:

>javac -O commandline.java

This should compile main benchmark driver and dependent files.

RUNNING THE BENCHMARKS

From the directory above this one, issue the command:

>java jnt.scimark2.commandline

or >java jnt.scimark2.commandline -large

to run the large problem size version. (Note that this one takes considerably longer to run.)

After a few minutes, the program should respond with the benchmark results, e.g.

>javac jnt.scimark2.commandline

SciMark 2.0a

Composite Score: 20.791595999749727
FFT (4096): 30.260047144878346
Jacobi SOR (100x100):   33.074935359763934
Monte Carlo (25000): 11.510791361970528
Sparse matmult (nz=25000), 10 iterations: 8.007507030681996
LU (100x100): 21.104699101453836

java.vendor: Sun Microsystems Inc.
java.version: 1.2
os.arch: x86
os.name: Windows NT
os.version: 4.0

One can send these results to "pozo@nist.gov".

Credits

https://math.nist.gov/scimark2/credits.html

Thank you to all who have contributed SciMark results.

SciMark 2.0 was developed by Roldan Pozo, and Bruce Miller at the National Institute of Standards and Technology (NIST). The goal was to better understand the JVM/JIT behavior of various Java platforms for numerical computing. The benchmark is similar in spirit to the Matlab and Java Linpack benchmarks, both of which have proven to be very useful.

The Java FFT kernel was based on components of the GNU Scientific Library (C code developed by Brian Gough).

Karin Remington developed the Perl scripts in the original SciMark to process the incoming result mail into a usable performance database, suitable for displaying in HTML.

SciMark 2.0 contains components from the Java Numerical Toolkit (JNT), an ongoing project at NIST to develop numerical libraries and tools for the Java community.

We would also like to acknowledge the Java Grande Forum and in particular participants of the Java Numerics Working Group for providing helpful discussions and valuable feedback. SciMark is also accessible from the Java Grande Forum Benchmark Suite home page.

License

As this software was developed as part of work done by the United States Government, it is not subject to copyright, and is in the public domain. We would, however, appreciate acknowledgements if this work is found useful. Note that according to GNU.org public domain is compatible with GPL.