SCL Cluster Cookbook
Introduction to Clusters

·   Cluster Cookbook Home   ·   Scalable Computing Lab   ·

Extraordinary improvements over the past few years in microprocessors, memory, buses, networks, and software have made possible the collection of groups of inexpensive personal computers and workstations that in concert have processing power rivaling supercomputers. Clusters of sixteen commodity microprocessor systems now achieve gigaflop-class performance at a total cost of under $50,000.

Sixteen nodes in a cluster
A cluster is a group of independent computer systems and thus forms a loosely-coupled multiprocessor system. A network is used to provide inter-processor communication. Applications that are distributed across the processors of the cluster use either message passing or network shared memory for communication. Because of the low prices of Intel-based PCs, most clusters are built using plain PCs with large memories and high-performance networks. Applications are often parallelized by using the MPI or PVM message-passing libraries for inter-processor communication.

Unfortunately, building a cluster is not a simple undertaking. The components making up the cluster should be tuned for the best performance on its primary applications. For example, an application may work better on a cluster with more processors but slower network technology than on a cluster with a very high bandwidth network but fewer processors. Also, commodity computer technology continues to change rapidly; an area where this is especially true is DRAM technology, which has gone from FPM to EDO to SDRAM in the past two years, and new, faster technologies are on the horizon.

As a result, there is no single standard recipe for building clusters (although we do give a sample recipe for a basic four-node cluster). In this cookbook, we will attempt to point out the major features of various current technologies, help you understand how the parts all fit together, and give hints for evaluating components before taking the "big plunge" and purchasing the entire system. At the Scalable Computing Laboratory, we have built several clusters based on Intel, SGI, and Digital Alpha CPUs and various network technologies. Each cluster has been custom built for use by different departments with different needs. Drawing from our experience, we will explain our methods in the hope that building your cluster can be as simple as possible.


Questions or comments? Send mail to ghelmer@scl.ameslab.gov
Copyright © 1997, 1998 All Rights Reserved by Scalable Computing Laboratory.
Disclaimer Notice
Maintained by ghelmer / Last updated on 06/10/98