The Aggregate Cluster Design Rules

If there is one unifying theme in all our cluster supercomputing work, it is that clusters need to be engineered to perform well on the codes you care about most. After many years of engineering, building, and using Linux PC clusters, we have finally been able to partially automate the engineering process. The tool we have built is not perfect, and you should treat its designs primarily as suggested starting points rather than as the design you should use, but it handles a surprisingly complex mix of issues very well -- mostly because it is really a specialized expert system that evaluates every possibility.

The tool is implemented as a CGI program accessed via a WWW form interface. To use it, you need to know a little about your application... and to tell our CGI about your design constraints. Of course, the fact that a particular part appears in the database, or is selected by the CGI, should not be considered an endorsement of that product by the author of the CGI (professor Hank Dietz) nor by the University of Kentucky; responsibility for evaluating the true fitness of any design falls entirely on the user of the tool.

A more detailed desrciption is available in our Cluster Monkey article entitled A Web-Based Tool for Optimized Cluster Design. A version of this tool is now available at http://cgi.aggregate.org/cgi-bin/cdr.cgi

The current bug/wish list is:

• A way to specify latency and bandwidth when a PC node is used as a router for a particular type of network hardware. This is needed to make the evaluation of toroidal hyper-mesh network designs (i.e., 1D ring, 2D wrap-around mesh, and 3D wrap-around mesh) more accurate.
• The current CGI assumes that each network port in a node takes 1 NIC and 1 PCI slot. We want to add support for specifying both ports built-into the motherboard and multi-port NICs.
• Support of "mixed bandwidth" network designs. The current evaluation of switched networks assumes that all links are of the same type. However, some 100Mb/s Ethernet switches have Gigabit Ethernet uplink ports, so a mix of 100Mb/s and Gigabit switches is possible. Although it would be nice to have to tool design mixed fabrics, they can be evaluated by the current version of the tool by specifying a switch part that has the characteristics of the entire fabric.

Source code is available from the "Beowulf Design Rules" project web page hosted by .

Publications

• William R. Dieter and Henry G. Dietz, A Web-Based Tool for Optimized Cluster Design, Cluster Monkey, January, 12, 2007.
• William R. Dieter and Henry G. Dietz, Automatic Exploration and Characterization of the Cluster Design Space , Technical Report TR-ECE-2005-04-25-01, University of Kentucky, April, 2005. BibTeX

Author Info

This program was originally built by Hank Dietz, but has been dramatically improved by Bill Dieter. Both are faculty in the Electrical and Computer Engineering Department at the University of Kentucky.

The only thing set in stone is our name.