PIC is a throughput-oriented site, where latency is not so important but final throughput is the critical point. This means that it is more important for us to finish as many jobs as we can, even if that means individuals jobs can take much more time to finish.
Thus, instead of having a huge supercomputer with low-latency links as the way of getting our computing power, our approach is having replication of many small commodity units that, when working together, achieve our goal.
For coordinating those units, we use a batch system. The software solution that was chosen is Torque/MAUI.
The hardware solution is currently based on HP: we currently have (as of 13/07/2009) 11 HP c7000 blade enclosures installed, full with 16 HP Blade BL460G1/G5 each. Each blade has 2 chips with 4 cores each, for a grand total of 1408 cores available to our batch system. The connectivity from-to the storage system has a bandwith of 30 Gbps, using 3 10GEthernet switches. This is one of the keys of a good utilization of those cpus that are demanding in terms of input/output.
