High-Performance Computing Cluster Apollo

Performance: 1.04 TFLOP/s (theoretical peak)
Manufactured by DELL in 2008
Let's see what it is made of...

Configuration: 50 compute nodes (200 processor cores), one master node, and one NFS server, connected by two 1 Gigabit Ethernet switches. Each compute node has dual 2.6 GHz Dual Core AMD Opteron™ processors, 4 GB of RAM, and a 160 GB local scratch disk. The master node offers 584 GB (RAID 5) user space, dual 2.4 GHz Dual Core AMD Opteron™ processors, and 4 GB of RAM. The NFS server provides 2 TB (RAID 5) cluster scratch disk, and utilizes two 1.6 GHz Quad Core Intel Xeon™ processors and 2 GB of RAM.

Software: Red Hat Linux, Platform OCS, Lava

Platform OCS

We use Platform OCS cluster software distribution. This provides for a straightforward installation and for uncomplicated maintenance. Here is a link to the Platform OCS User Guide.

Lava

For job scheduling and cluster resource management, we use Platform Lava. It is based on Platform LSF. It is easy to use and allows students to get familiar with one of the popular work load management systems. Please follow this link for a description of how to use Lava.

Architecture and Purpose

As mentioned above, the main components of cluster Apollo are the master node (the front end node), 50 compute nodes, one NFS server, and the 1Gb Ethernet network provided by 2 switches.

The front-end (master) node is connected to the CCB network and used for cluster user login and directories. It also runs the job scheduling and cluster management software. One can perform all work on the cluster form the master node. The front-end also provides a Web based interface, which is accessible from the CCB network using this link. This website offers current cluster resources allocation information, Lava GUI, and cluster documentation.

The NFS server provides the network-attached storage to all compute nodes, mounted under /apollo-io/scratch. Please do not run any computing or data processing on this node.

All 50 compute nodes are used for running jobs as scheduled by Lava batch system. The local /scratch partition on the compute nodes is used for job data. Please remove the data after the job is done.

Though the 1Gb/s network is supported by 2 Ethernet switches, all 50 nodes build a single pool of computing resources available to the job scheduling system. The stacking link between the switches provides data bandwidth of 12 Gb/s. This cluster is designed for loosely-coupled calculations, which require only a moderate data bandwidth between compute nodes.