Grendel: The Clemson Beowulf Workstation

Grendel, part of the Parallel Architecture Research Laboratory, is an experimental parallel computer built from commodity components. Specifically Grendel is a Pentium-based pile-of-PCs of 18 machines, each with the following: The machines are tied together with two networks. The first is a bus network using a stack of AsanteFast 100 Hubs. The second is a full-duplex switched network using a Bay Networks 28115/ADV fast ethernet switch. Two nodes are used for interacting with the system, while the other 16 are dedicated as compute and I/O servers.

This is only one example of a Beowulf Workstation, which is a concept developed at NASA. Really the concept includes not only off-the-shelf hardware but also the use of freely available operating systems such as Linux, message passing software such as PVM and MPI, and other software often contributed by Beowulf users. More information on the Beowulf concept and links to other sites with these machines are available from the CESDIS Beowulf Page and the BEOWULF Research Page.

One focus of work on this machine is in parallel I/O. Grendel serves as the testbed system for the Parallel Virtual File System Project. The goal of this project is to find parallel I/O solutions for pile-of-PC architectures and to make parallel I/O available to a larger group of users. An example of the benefits of parallel I/O on a Beowulf Workstation can be found in ``Support for Parallel Out of Core Applications on Beowulf Workstations'', to appear in the proceedings of the 1998 IEEE Aerospace Conference. This paper details the performance of a Gauss-Seidel iterative solver on a Beowulf workstation using both an in-core and an out of core implementation.

The Clemson Beowulf Workstation is also part of the Clemson Regional Data Center, where it is used to parallelize key RDC applications for the Beowulf architecture.

For more information on the Clemson Beowulf contact Walt Ligon at walt@clemson.edu or visit the Parallel Architecture Research Lab home page.


Part of the Parallel Architecture Research Lab