New Cluster for Oxford Supercomputing Centre
The Oxford Supercomputing Centre (OSC) is pleased to announce the availability of it's new cluster, “SAL”. SAL is an SGI ICE cluster of the same type as the older OSC cluster 'HAL', but using the new Intel “Nehalem” generation of processors.
The system arrived on the 22nd of February and was running test jobs within a few days. SAL can be used in exactly the same way as the HAL system so there is generally no need to rebuild code which has previously been compiled on HAL. The two systems have the following specifications:
| Component | HAL | SAL |
|---|---|---|
| Processors | 2.8GHz Quad Core Harpertown | 2.53GHz Quad Core Nehalem |
| Memory per core | 2GB | 3GB |
| Cores per node | 8 | 8 |
| Number of compute cores | 512 | 640 |
| Scratch space | 12TB | 12TB |
Early benchmarks show that most applications can expect a 20-30% percore performance boost over HAL even though the processors' clock speed is lower. In some cases (where memory is a bottleneck) SAL is even faster as its measured memory bandwidth is more than double that of any other OSC system. However, applications which do not use main memory very much may still get their best performance from HAL. We therefore strongly recommend all users try SAL to see what performance increase they can expect, especially if they have been restricted to using QUEEG by memory limitations.
To submit jobs to SAL, please log onto sal.oerc.ox.ac.uk or sal.osc.ox.ac.uk.
There are two other features of SAL which may be of interest to more advanced users. Firstly, hyper threading is enabled giving 8 real and another 8 virtual cores per node. While we strongly recommend using no more than 8 MPI processes per node unless they are very light on memory usage, it may sometimes be possible to run up to 16 processes per node at reasonable efficiency. This would be of particular use to hybrid MPI/shared memory applications. If you'd like to try this, please talk to the OSC staff first as we'd be interested to see how theory matches practice and 16 memory heavy MPI processes are likely to crash compute nodes!
Secondly, HAL and SAL are physically joined together. Users who need to run a small number of very large jobs (>512 cores) and which are well load balanced are welcome to talk to the OSC staff in order to make special arrangements.
We will be updating our website over the next few days to include SAL's arrival, but please feel free to try the new machine out using the instructions already provided for HAL. As always, do report any issues to support@osc.ox.ac.uk.