To meet growing demand for delivering high-speed supercomputer services to researchers across Japan, Kyoto University has deployed a cluster supercomputer based on Intel’s Xeon processor E5 family and featuring a 1,202-Socket Subsystem and high-memory-capacity Subsystem with 1.5TB per Node.
As one of Japan’s national centers for shared access to IT infrastructure, the university’s Academic Center for Computing and Media Studies also conducts research into the building of advanced IT platforms for the future, explains Prof. Hiroshi Nakashima, Director of ACCMS.
“To keep up with other academic institutions around the world, it is important that we progress in line with the latest developments. In order to do this, we configure systems as Linux* clusters, an architecture that is in widespread use internationally.”
After comparing and testing machines from a number of vendors, they configured a large cluster system using an HPC server fitted with the Intel Xeon processor E5 family.
One MPP, Two Clusters = 553 Tflops
The new HPC includes one massively parallel processor (MPP) system and two cluster systems, each comprising an InfiniBand* network and HPC server fitted with the Intel Xeon processor E5 family. The peak computational performance of the overall system is 553.9 TFlops.
The Laurel subsystem has a high degree of compatibility with PC clusters, comprising 601 nodes with 64 GB of memory and 16 cores per node. It has a peak computational performance of 242.5 TFlops and total memory capacity of 38 TB (Figure 3). The Cinnamon subsystem has 16 nodes with 1.5 TB of memory and 32 cores per node.
Despite the small number of nodes, the large amount of memory per node means it will be used for applications that demand a large memory capacity. It has a theoretical peak computational performance of 10.6 TFlops and total memory capacity of 24 TB .
High Speed Analysis at Low Cost
While the Cinnamon subsystem, with its large memory capacity, has only improved node performance by about 20 to 30 percent, power consumption has been cut to one-tenth that of the system it is replacing. Also, the ability of the entire system to fit into a single rack means it takes up only one tenth as much space as the previous system.
The new system represents 7.9 times the computational performance, 6.1 times the memory capacity, and 5.7 times the overall physical capacity of its predecessor. The majority of use comes from science and engineering research, with examples of large simulations on which the center has collaborated including simulations of the cycle of major earthquake events and particle simulations of plasma environments. Applications include large scientific and technical computations, computational chemistry, structural analysis, statistical processing, and visualization.
Flexibility is key. Prof. Nakashima explains: “The supercomputers made available by ACCMS can be used by all sorts of different research institutions in a variety of fields and disciplines. This means, rather than being mission-oriented, our service must maintain a highly versatile approach.”
Clear Benefits to Researchers
The improvement in underlying performance means users can obtain the results of even large and complex calculations quickly and cost-efficiently. In an academic context, the faster speed will prove valuable because it allows calculations to be executed for more parameters than would otherwise be possible in the limited time available. The new system also has benefits for ACCMS in its role as a service provider.
Under current operating practices in which usage fees are calculated based on the amount of power that users consume, a major benefit is the near six-fold improvement in power efficiency, explains Professor Nakashima, “which means that we can provide users with roughly six times as much computing capacity within the allocated budget.”
More growth in 2014
ACCMS has already decided to install an additional new supercomputer fitted with the next generation of Intel Xeon processors in 2014. The new subsystem is expected to research institutes throughout Japan using the Intel® Xeon® processor E5 family have a peak computational performance of 400 TFlops. Combined with the existing system, this will result in a supercomputer with performance approaching the 1 PFlops range.