IBM's BlueGene technology became available in 2004/2005, and is now the leading system for capability computing applications. A key feature of the BlueGene architecture is its scalability, low power consumption, and good price-performance ratio.
Jülich was one of the early adopters of BlueGene technology: in 2005, Jülich started testing a single BlueGene/L rack with 2,048 processors. It soon became obvious that many applications could be ported to the Blue Gene architecture, and since the BlueGene architecture is well balanced in terms of processor speed, memory latency and network performance, many of these can be success fully scaled up to large numbers of processors. In January 2006, Jüich therefore upgraded its BlueGene/L system to eight racks with a total 16,384 processors using funding from Germany's Helmholtz Association.
The eight-rack system has been in operation successfully for two years now, and about 30 selected research projects use it to run their applications -- this includes lattice QCD simulations that probe the complicated interactions between the quarks and gluons that make up atomic nuclei.
This large user demand and its strategy to strengthen leadership-class computing caused Forschungszentrum Jülich to procure a powerful next-generation BlueGene system, and in October 2007, a 16-rack BlueGene/P system with 65,536 processors was installed, financed mainly by the Helmholtz Association and the State of North-Rhine Westphalia.
With a theoretical peak performance of 222.8 TFlop/s and a measured computing power of 167.3 TFlop/s, Jülich's Blue Gene/P - dubbed JUGENE - was ranked second in the TOP500 list of the fastest computers in the world as of November 2007. The only computer ahead of JUGENE being used for military purposes, JUGENE can lay claim to being the world's fastest civilian computer.
The main improvements of the BlueGene/P over its BlueGene/L predecessor concern the processors and networks, while the principal design remained unchanged. The new key features of BlueGene/P are: four PowerPC 450 processors combined in a four-way SMP chip, allowing a hybrid programming model with MPI and OpenMP with up to four threads per node, and a DMA-capable network interface for increased network performance and reduced processor load during message handling.
The available memory per processor has been doubled and the external I/O network has been upgraded to 10 Gigabit Ethernet. These improvements are also reflected in the application performance: Lattice QCD simulations see a speedup of about 20% on BlueGene/P compared to BlueGene/L. Furthermore, the increased memory of 2 GB per node will allow new, particularly memory-hungry, applications to be run on BlueGene/P.
With JUGENE Forschungszentrum Jülich has taken the next step towards petascale computing and has strengthened Germany's international leadership position in scientific supercomputing. (And let's be honest, it has also given computational scientists like myself [running code on JUGENE as I write this] a beautiful new to
This article incorporates information taken from a press release of Forschungszentrum Jülich.
