Thunder: the world's next fastest Linux supercomputer

LAWRENCE LIVERMORE National Laboratory (LLNL) is one of the key high-performance computing sites in the States, with lotsa $$$ poured at advanced architectures research, numerical methods and computer science.

They have developed a massive multiplatform clustered supercomputing, which for now revolves around various Intel CPUs - after all, they're not far from Satan Clara - and the Quadrics QsNet high-speed interconnect from Bristol, UK.

Their newest baby, called Thunder, is capable of supporting a complex workload consisting of medium (1,024-2,048) and large (2,048-4,004) MPI task count parallel jobs for unclassified Multiprogrammatic & Institutional Computing (M&IC) simulations - in essence, go up as close to the classified defence - level performance as it is possible.

How to get there? They combine 1,002 quad-Madison 1.4 GHz Intel Tiger4 boxes (each 1.4 GHz Madison is 5.6 GFLOPs peak, so times 4,008 for total Rpeak of close to 23 TFLOPs - I guess Rmax could go as high as 19 GFLOPs when tuned well) with federated-switch interconnect based on QsNet II (Elan4) - the brand new version of Quadrics interconnect, with some record-breaking performance numbers for any kind of cluster.

Not to mention that, besides message passing, this interconnect lets them use Cray-style "shmem" distributed shared memory approach as well, letting the user see the whole memory of the cluster as one logical address space, without need to replicate the data like in message passing.

A lot more there - the revolutionary Lustre cluster file system, ultrafast federated Gigabit Ethernet backbone as well, connection to two more large existing clusters, plus a visualisation cluster - see the diagrams and enjoy looking at taxpayer's dollars at work!

The Thunder thingie will be deployed this November - so better queue up for the time... µ