Page 109 - 49A Field Guide to Genetic Programming

10.5 Geographically Distributed GP 95

Figure 10.3: A globally distributed GP system (Langdon, 2005a). The
server is the centre of the star architecture, with the lines connecting it
to users around the world. The users evolved snowflake patterns using a
continuously evolving L-System, and their (subjective) preferences provided
the fitness measure used to drive the system.

While many have looked enviously at Koza’s 1000 node Beowulf cluster
(Sterling, 1998) and other supercomputer realisations of GP (Bennett, Koza,
Shipman, and Stiffelman, 1999; Juille and Pollack, 1996), a supercomputer is
often not necessary. Many businesses and research centres leave computers
permanently switched on. During the night their computational resources
tend to be wasted. This computing power can easily and efficiently be
used to execute distributed GP runs overnight. Typically, GP does not
demand a high performance bus to interconnect the compute nodes, and
so existing office Ethernet networks are often sufficient. While parallel GP
systems can be implemented using MPI (Walker, 2001) or PVM (Fernandez,
Sanchez, Tomassini, and Gomez, 1999), the use of such tools is not necessary:
simple Unix commands and port-to-port HTTP is sufficient (Poli, Page,
and Langdon, 1999). The population can be split and stored on modest
computers. With only infrequent interchange of parts of the population
or fitness values little bandwidth is needed. Indeed a global population
spread via the Internet (Chong and Langdon, 1999; Draves, 2006; Klein and
Spector, 2007; Langdon, 2005a), `a la seti@home, is perfectly feasible (see
Figure 10.3).
Other parallel GPs include (Cheang, Leung, and Lee, 2006; Folino, Piz-
zuti, and Spezzano, 2003; Gustafson and Burke, 2006; Klein and Spector,
2007; Tanev, Uozumi, and Akhmetov, 2004).