ClusterFsck Page

This page covers our work on cheapo Linux clustering - briefly, we built an eight-node Linux cluster for parallel processing research. The cluster has been retired (see below), but I've left the page up since it still gets read now and then.

Background and Motivation

In late 1998 or so, Christopher Woody and I decided to build a Beowolf-style cluster based on Linux. However, being grad students, we were broke.

However, this was in Albuquerque, and every week or two, there was an auction of surplus hardware from Los Alamos and Sandia labs. If you were patient, you could get lots of old hardware really cheaply, e.g. a 486 PC for under $10 each.

It's pretty easy to build a cluster when you can call the harware vendor and get N identical machines. It's a lot more difficult when every single machine is different, and most are flaky...


The goal was to build a d=3 (eight node) hypercube for $100USD or less.

We failed.

We spent $105USD.

We had 8 compute nodes, one router, and one server / console. We wrote code to queue up POVRAY raytracing jobs, and got excellent performance - one the close order of half of a Pentium II's performance. Check out the pictures for output from the parallelized renderer.

We were also using spare compute cycles on the RC5 project, with the DX2/66 nodes getting ~65k keys/sec, the DX50 ~50k, and the rest somewhere in between.

The impetus was Dr. David Bader's excellent Parallel Algorithms class, and the local availability of super-cheap 486 machines. The basic hardware is procured from the bi-weekly local government surplus auctions, see the Bently's pages

Hardware configuration

Each node was a DX2/66, (actually, we had one DX50 and one DX4/75, the rest were 66) with at least 16MB memory, local disk for the OS (We were using the excellent SuSE Linux distribution), and TP ethernet to the hub. We had two additional nodes: One minimal Linux (386SX/25, 8MB, dual 3c509) box used as a router, and one, slightly faster DX4/75, used as a console and server for the other nodes. The topology was currently a bus; we never did manage to get our hands on an Ethernet switch to try more communication-intensive computing.


This link is a bzip'd tar file containing the ApplixWare 4 presentation and poster materials that we used to create our final presentation.

Status Update June, 1999

The cluster has been passed on - an era has ended. ;) Bug Ken Segura if you want more information as to what it's up to now.


We (myself and Christopher P. Woody) gratefully acknowledge the invaluable assistance of Frank Mercer of the EECE department, who has contributed vast amounts of hardware and assistance. Without him this project would never have been possible.

Other significant donors include Mark Walker (EECE systems admin), ourselves, and Barry Keeney of Los Alamos National Labs. (The pictures were all taken on his Kodak digital camera)


I've dug up the pictures taken on the Sony Mavica camera, and these are the best of the bunch.

The two rendered pictures are output from the ClusterFsck, using our software to distribute POVRAY jobs across nodes. The sombrero is an included example, and the soda glass is a creation of Timothy Eyring, a friend of ours.


Is it just me, or does Chris Woody resemble George Michael from the early eighties?

58210 bytes
480 x 640
83175 bytes
480 x 640
58284 bytes
640 x 480
55944 bytes
480 x 640
42577 bytes
640 x 480
63323 bytes
640 x 480
57861 bytes
480 x 640
55904 bytes
480 x 640
65373 bytes
480 x 640
58166 bytes
480 x 640
148216 bytes
640 x 480
91233 bytes
480 x 640
257976 bytes
800 x 600
116173 bytes
1024 x 1024
68180 bytes
480 x 640
55075 bytes
480 x 640
53299 bytes
480 x 640
53171 bytes
640 x 480
69725 bytes
480 x 640


  • Back to home page
  • Generated by HTMLThumbnail 1.1.4