Abstract

Existing Beowulf clusters are normally limited to the actual number of compute nodes physically connected to the cluster’s network switch. At some point, it may become necessary to extend the size of the cluster beyond the capacity of the switch. By using existing computers on a campus network, one can extend the computing power of the cluster by including these machines during times they would normally be idle. We have done a survey of methods for extending clusters and devised a parallel solution to a computationally complex all-possible­regressions algorithm and tested it by using idle computing power on the campus network.

First posted 2003