Optimizing the parSOM neural network implementation for 
  data mining with distributed memory systems and cluster computing

          Philipp Tomsich, Andreas Rauber and Dieter Merkl,
               Department of Software Technology, 
                Vienna University of Technology,
                    Favoritenstr. 9 - 11 /188, 
                     A--1040 Vienna, Austria,

                   www.ifs.tuwien.ac.at/~phil
                   www.ifs.tuwien.ac.at/~andi
                   www.ifs.tuwien.ac.at/~dieter


Abstract:
The self-organizing map is a prominent unsupervised neural network
model which lends itself to the analysis of high-dimensional input
data and data mining applications.  However, the high execution times
required to train the map put a limit to its application in many
high-performance data analysis application domains, where either very
large datasets are encountered and/or interactive response times are
required. We demonstrate the merits of self-organizing maps in the
field of text classification, which forms a prominent application
domain for high-performance computing.

In this paper we discuss the parsom implementation, a software-based
parallel implementation of the self-organizing map, and its
optimization for the analysis of high-dimensional input data using
distributed memory systems and clusters.  The original parsom
algorithm scales very well in a parallel execution environment with
low communication latencies and exploits parallelism to cope with
memory latencies.  However it suffers from poor scalability on
distributed memory computers. We present optimizations to further
decouple the subprocesses, simplify the communication model and
improve the portability of the system.

----

In: Proceedings of DEXA-Workshop on Web-based Information Visualization
(WebVis~2000), 4. - 8. Sept. 2000, Greenwich, UK;
Tjoa, {A M.}, Wagner, {R.R.}, and Al-Zobaidie, A. (Eds.), 
pp. 615 -- 619, IEEE Computer Society Press, 2000.