RESEARCH/Open Supercomputer Specifications (Detailed)

Complete Description of the T2K Open Supercomputer Specifications

  • A T2K Open Supercomputer must consist of shared memory nodes of sixteen or more 64-bit IA32 CPU cores.

  • The aggregated memory bandwidth between the node memory and CPU cores must be 40 GBtyte/s or larger.

  • Each node memory must have SECDED (single-error-correction/double-error-detection) error correction/detection mechanism.

  • Each node must have a link or an even number of links whose (aggregated) bandwidth is 5 GByte/s or larger. If the data transfer through the link is performed with bit encoding such as 8B/10B encoding in its physical layer, the bandwidth should not be the physical bit transfer rate but should be the data transfer rate calculated taking the encoding overhead into account.

  • The minimum round-trip latency of a point-to-point MPI communication through the inter-node link(s) must be 8.5 μsec or less. The maximum data transfer bandwidth of a point-to-point MPI communication through the inter-node link(s) must be 4GByte/s or larger.

  • Each node must have multiple hard-disk drives whose total capacity is 250 GByte or larger. The hard-disk drives must be capable of RAID-1 operation without operating system functions.

  • Each node must have a hardware management interface compliant with IPMI 2.0 or later.

  • The operating system running on each node must be Red Hat Enterprise Linux 4, SuSE Linux Enterprise 10, or a Linux which is proven to be equivalent to or more advantaged than the either of them. The operating system should support 64-bit processor architecture and applications.

  • Compilers of Fortran, C and C++ must be provided and must have auto-parallelization functions and OpenMP (2.5 or later) API.

  • A Java system with JIT compilation function must be provided.

  • A communication library compliant with MPI 1.2 or later must be provided.

  • Mathematical computation libraries BLAS, LAPACK and ScaLAPACK must be provided.

  • The basic library provided with the operating system, the MPI library and the mathematical computation libraries above must be freely usable without any limitation of the number of execution instances and programs with the libraries must run on any nodes on the system.