Compact Cluster Computing

Arnie Stillman

Physics Department, Brookhaven National Laboratory

 

One way to make inexpensive fast computing clusters is to increase network bandwidth in the cluster. Implementing a parallel bus for the network physical layer can increase bandwidth to its maximum, and the bus itself increases the cluster density dramatically. In a typical cluster, the universal communication protocol is TCP/IP, and the physical layer is generally ethernet. I am proposing the bus and a bus bridge to generate parallel-TCP/IP ethernet as a method of constructing a cluster in a single rack. This cluster has the potential to contain on the order of ten to fifteen processors in a single crate, all sharing a network with maximum bandwidth. The processors themselves are small single board computers, hence the name "Compact Cluster Computers." Since this is a cluster on ethernet, scaleability is no more difficult than any other cluster implementation. Such clusters are not optimal for intensive numerical computations, but they can be very efficient in servers that run multi-threaded applications or thousands of duplicate processes.