Research

Main research topics

Large scale supercomputing / high performance computing

We are researching in various aspects of high performance computing technology including system architecture, interconnection network, system software, applications and tuning, based on the collaborative work in Center for Computational Sciences with its resources such as T2K-Tsukuba, FIRST and recent HA-PACS. The most important work recently is the system design and construction of direct communication network system for a large number of accelerators not relying on CPU for large scale parallel processing. The target applications include Lattice QCD, Astrophysics, Geo-Science (Climate) and Bioscience.

High performance computing with GPGPU

We are researching on the GPGPU technology to drive the computation power in various application area especially on how to reduce the overhead of communication among GPUs not just within the computation node but over large scale parallel systems. We target both ways using NVIDIA-CUDA directly and using GPU-ready compilers such as PGI or CASP.

Large scale GPU cluster HA-PACS

HA-PACS project is started from 2011 to make a large scale GPU cluster with commodity technology as well as the research and development on a dedicated interconnection network technology for ultra low latency communication among GPUs over computation nodes. Current HA-PACS base cluster consists of 268 computation nodes where each node is equipped with two sockets of Intel E5 (SandyBridge?) 8 core processors and four of NVIDIA M2090 GPUs. The interconnection network is dual-rail Infiniband QDR with up to 8GB/s of bandwidth. The peak performance of single node reaches to 3 TFLOPS (320 GFLOPS by CPUs and 2.7 TFLOPS by GPUs). The total peak performance of base cluster of HA-PACS reaches to 802 TFLOPS. The system will be extended with new TCA (Tightly Coupled Accelerators) concept and latest GPUs on 2013 with additional hundreds of TFLOPS performance. We are expanding the research topics on HA-PACS with large scale parallel applications, TCA architecture development and parallel language based on PGAS concept and accelerator offloading named XMP-dev (XcalableMP accelerator DEVice extension).

Publications

English Papers


トップ   編集 凍結解除 差分 バックアップ 添付 複製 名前変更 リロード   新規 一覧 単語検索 最終更新   ヘルプ   最終更新のRSS
Last-modified: 2016-02-22 (月) 11:38:24 (1332d)