Tera-10, the new supercomputer of the Commissariat à l' Énergie Atomique (CEA), is a colossal machine, 602 nodes connected with three rails of Quadrics' high-performance QsNetII network. Designed by Bull, Tera-10 integrates 544 NovaScale 6160
compute servers, each with eight Intel dual-core Montecito processors,
with an I/O system that can store in excess of 1 Petabyte of data and
deliver it to the nodes at least 100Gbytes/sec. The compute cores share
27 terabytes of memory.
The system is called Tera-10 as it is designed to sustain at least 10
Teraflop/s on the CEA's application. Linpack performance is excess of
60 Teraflops/s, making Tera-10 the most powerful supercomputer in
Europe. Tera-10 runs the Bull HPC software stack, this includes the
BAS4 Linux kernel and distribution, the Lustre parallel file system, the Quadrics Resource Management System (RMS) and the Bull/Quadrics MPI.
Performance
- 60 Teraflops of computing power
- 27 Terabytes of memory
- 1 Petabytes of disc space with a throughput of 100GB/s
- 544 Bull Novascale compute servers
- 3 rails of Quadrics QsNetII interconnect
- 54 NovaScale I/0 servers
- Quadrics parallel fibre links
- Bull Linux kernel and distribution
- Quadrics resource management system
- Quadrics/Bull MPI
The impressive results are achieved through a combination of remarkable advanced features:
System Spec and Performance
Its main application is the simulation of atomic experiences and the maintenance of the French nuclear defence force, using the results of true nuclear tests (between 1956 and 1995, most of them in the French Nuclear Test Plant in the Pacific) and the new results obtained from the LMJ (Laser Mega-Joule) built in continental France.
This is the end of a second generation of evolutions, the next generation will reach about 100 TeraFLOPS (expected in 2009) using processors with more internal cores and a new central scheduling system allowing asymetric operation on a variable number of processors, for easier upgrades, lower maintenance cost, and more experiences requiring different computing scales, without having to rebuild the whole cluster. This new system should be renamed Tera-100.
Sites 