MUNCC
MUNCC stands for MUltiple Node Computing Cluster. It is a small Beowulf cluster that was first built in 2000. Since then, it has been revised several times, and we now refer to it as MUNCC3. While it was primarily developed for use in Molecular Dynamics simulations, it actually extends beyond the scope of that project as a lesson in clusters, Linux, and network administration.
We wanted a computer that could run our simulations somewhat faster than the average desktop we have lying around at school, plus the ability to add nodes in the future was flexability that we wanted. The nodes were salvaged after the school upgraded to laptops. On their own, the nodes aren't much, put as a group we noticed a significant performance increase.
Here is an overview of the system:
Master Node (Garion)
Garion runs all of the job control systems. Specifically, I installed Torque with MPI and Maui. It also has a hard drive that it shares with all of the nodes. This is where we host shared programs like NAMD, GROMACS, etc. Users should only ever interact with this computer, either trough the GNOME GUI terminal or SSH.
Services Node (Durnik)
Durnik does most of the services work. This node runs the firewal and acts as the only external facing computer. All SSH requests are silently routed through Durnik to Garion. It also runs an LDAP user database, Bind9 DNS, and DHCP for al of the other nodes,
Worker Nodes
The worker nodes all run on similar hardware. We currently have 8, but the system is designed to be able to add more with just a hard drve copy and a line added to the DHCP file. The flexibility in increasing performance was required in case we ever need to speed up long simulations. Torque passes all MPI enabled programs through to the worker nodes, effectivly splitting the work between all 8 nodes. I have tested the performance increase with each added node, as seen below.
Here is part of a presentation I gave on MUNCC. The last slide has the performance gained graph.
[UPDATE] This is another presentation I recently gave about how we can utilize Harmony Searches for automatic refinement, GPU for extreme performance gains, and Amazon EC2 for cheap HPC clusters: