UVic Computing  |  UVic Home
  

The Mercury Cluster

Mercury is a Linux cluster geared to serial jobs. It gets its power from 210 IBM blade servers in 15 BladeCenter chassis, each with dual Intel or AMD processors ranging in speed from 2.4GHz to 3.4GHz. This computing power is fronted by a pool of interactive login nodes that run a superset of software available on the compute nodes, augmented by compilers, libraries, editors, graphic and other tools.

Cluster Networking

The cluster's internal networking is based on Gigabit Ethernet. Each blade has a GigE connection to the chassis, which in turn has two GigE connections to the switch, channel-bonded to give, effectively, a single two-GigE connection. This same bonded connection exists between the head node and the switch.

The head nodes also connect to Storage through Gigabit Ethernet, and mounts research filesystems owned by various groups. These filesystems are in turn re-exported to the blades.

Cluster Scheduling and Management

Mercury's power derives from its ability to farm computing work out to its blades. This is accomplished via two open-source software packages: the Portable Batch System, and the Maui Scheduler. An additional package, the Extreme Cluster Administration Toolkit (xCAT), aids in the management of software on the blades.

PBS manages Mercury's job queues, tracks jobs, and retains information on resources available on each node and the node's availability.

The Maui scheduler coordinates when and where jobs are run. It examines jobs in the various queues, applies various factors and algorithms to determine their priority, and requests for PBS to dispatch the jobs accordingly.

Getting 210 independent machines up and running, updated to the latest software revisions and configured appropriately would be a nightmare without xCAT, which employs KickStart and PXEBoot (Pre-Execution Environment) to distribute software among the blades. With xCAT, blades can be wiped clean and issued a fresh software install with two commands.

Grid Enablement

The Mercury cluster is connected to Grid Canada's production grid via the Globus Toolkit. The Grid is a relatively new development in heavy computing that could be described as a wide-area cluster.

You can read more about the Grid at Globus.