eMinerals Lake clusters

Use of Globus

Three replica clusters are located at Bath, Cambridge and UCL (Bruin et al, in preparation), and are each called Lake. Each cluster has one master node and 16 slave nodes, all with Intel Pentium 4 processors running at 2.8 GHz, and with 2 GB RAM per processor. The nodes have Gigabit ethernet interconnections. Each master node acts as the job manager, supporting both PBS and MPI. Each master node also hosts a data vault for the Storage Resource Broker (see below), and acts as a Globus Gatekeeper. At the present time, the clusters in Cambridge and UCL run v2.4.3 of the Globus Toolkit, and the cluster in Bath runs v3.2; we are planning to soon update all clusters to v3.2. A fourth Linux cluster, based in Cambridge and called Pond, has 40 nodes with Intel Pentium 4 processors running at 1.7 GHz and with 512 MB RAM per processor. At the time of writing, we are also incorporating a small cluster of Apple G5 Xserve dual-processor nodes, each having 8 GB RAM, and called Lagoon. The comparison between these different clusters highlights the fact that a minigrid structure contains provision for studies that have different requirements, some having high memory requirements and some having high-throughput requirements. The master nodes on each cluster also act as the Globus gatekeepers to other resources on their local networks; in the case of UCL and Cambridge these nodes are also the gatekeepers for access to the Condor pools described below.

At the time of writing, the clusters each run Mandrake Linux 9.2 and Oscar 2.3.1. Oscar provides all the necessary cluster installation tools, together with tools for all the basic operations such as job scheduling and basic monitoring via the Portable Batch System (PBS) and Maui job scheduler.

Most of the community of users use code written in Fortran, and so the clusters were provided with several Fortran compilers and MPI (Message Passing Interface) libraries (both MPICH and LAM implementations). This was driven by the fact that the experience of our users is that the performance of different codes is affected by the choice of compiler and MPI implementation. Finally, we also installed various Maths Libraries including ScaLAPAC, BLACS, etc., each of which was installed with their standard configuration.

Useful links

http://www.mandrakesoft.com
http://oscar.sourceforge.net
http://www.openpbs.org
http://www.clusterresources.com/products/maui

General references

Papers that describe the eMinerals science areas are:

---