Little-Fe: A Portable HPC Cluster for Computational Science Education

Little-Fe/PowerPC setup on top of the Pelican travel case at the Milburn Hotel in New York City. March, 2006.

Many institutions and teaching environments do not have access to parallel platforms for parallel and distributed computing education. Teaching key concepts such as speedup, efficiency, and load balancing are much more effectively done on a parallel platform. Little-Fe is a complete 4 to 8 node Beowulf style portable computational cluster. Little-Fe weighs less than 50 pounds, easily and safely travels via checked baggage on the airlines, and sets-up in 10 minutes wherever there is a 110V outlet and a wall to project an image on. By leveraging the Bootable Cluster CD project, and its associated curriculum modules, Little-Fe makes it possible to have a powerful ready-to-run computational science and HPC educational platform for under $2,500.

• N node computational cluster (4 node PowerPC)
• Portable
• Inexpensive (about $2,500 USD)
• Low friction, uses the Bootable Cluster CD distribution and a single system image

• High performance computing and computational science education
• Almost all of the available HPC cycles go to research, teaching needs are a distant second

• Need predictable, consistent access to HPC resources to teach effectively:
  • Immediate feedback, compute it now
  • Realistic science, no more computing Pi with a simulated dart-board

• Not reasonable to teach HPC/CS in a uniprocessor environment:
  • Decomposition - domain, functional
  • Speedup
  • Efficiency
  • Parallel models - message passing, shared memory, processing migration, etc.

4 - Pegasos II motherboards, 1GHz PPC processor, 256MB RAM
4 - Nano PW120 power supplies
2 - IBM Travelstar 40GB laptop form-factor disk drives
1 - CD-RW/DVD-RW optical drive
2 - D-Link 5 port gigabit switches
10 - CAT-5E Ethernet cables
1 - Linksys WRT54G access point/router
1 - Meanwell S320-12 320 Watt 110VAC-12VDC power supply
1 - Kattrel 6U subrack and rails
2 - RadioShack 12VDC-7.5VDC power supplies
1 - Aluminum channel and mounting hardware
1 - Pelican 1610 traveling case

Configure and test the effect of traffic partitioning on network throughput given an MPI workload.

Configure 2nd gigabit network interface and test the effect of channel bonding on network throughput for MPI workloads.

Quantifying and reducing TCP/IP network stack latency in the 2.6.x Linux kernel. See http://cluster.earlham.edu

Bootable Cluster CD Liberation testing. See http://cluster.earlham.edu

SIAM's Computational Science and Engineering - San Francisco, February, 2006.

ACM's Principles and Practices of Parallel Programming - New York, March, 2006.

Kean University - New Jersey, March, 2006.

National Science Foundation's headquarters tour - Arlington, VA, March, 2006.

National Computational Science Institute's Parallel and Distributed Computing workshop - Houston, TX, May, 2006.

TeraGrid '06 - Indianapolis, IN, June, 2006.

National Institute for Technology and Liberal Education's Introduction to Parallel and Cluster Computing workshop - Georgetown, TX, June, 2006.

SuperComputing education program development workshop. July, 2006.

EuroPar 2006 - Dresden, Germany, September, 2006.

National Science Foundation's headquarters tour - Arlington, VA, October, 2006.

Montgomery Blair High School presentation - Montgomery, MD, October, 2006.

Joint Educational Facilities presentation - Washington, D.C., October, 2006.

SC06 - Tampa, FL, November, 2006.

• Paul Gray - University of Northern Iowa
• Dave Joiner - Kean University
• Thomas Murphy - Contra Costa College
• Charles Peck, Tobias McNulty, Skylar Thompson, Kevin Hunter, Alex Lemann, and Josh McCoy - Earlham College