RESEARCH/Open Supercomputer Specifications
How Opened? - three objectives to open -
We named our specifications for new supercomputers Open Supercomputer Specifications aiming at making our supercomputers "open" in the following three aspects.
Open Hardware Architecture
The specifications clearly require that the supercomputers have to be built by commodity devices and technologies, rather than those dedicated only to high-performance computing.For example, their shared memory node must consists of processors of 64bit x86 architecture which is obviously the mainstreamer of current and near future processor technology, and DDR2 memories also commonly used in servers and high-end PCs.As for the interconnection between nodes, our requirements can be satisfied by a bundle of links of Inifiniband x4 DDR or Myrinet 10G both of which are leading interconnection technologies for commodity PC clusters.Since these devices and technologies are the most advanced and cost effective in the current IT market, the resulting architecture should be the best solution in the range of our procurement budget.
Open Software Stack
We also clearly state that the software stack should be built on top of the open-source operating system Linux. The other major components of the software stack, OpenMP based parallelizing compilers, MPI communication library and Grid middleware and tools, also have open-source solutions commonly used in the PC cluster community.Therefore, it should be easy for PC cluster users to port their programs to the open supercomputers having the open software stack.Moreover, the openness will give us the high portability of software components produced by advance research work of worldwide including ourselves.
Open to User's Needs
The supercomputers should be much larger and faster than the computers which our users are daily using but have the hardware and software architecture same as the user's own daily environment.Therefore, there are no barriers to use our supercomputers not only for users with numerical applications but also for those who have never used supercomputers for their non-numerical applications.That is, our supercomputers welcome emerging supercomputational scientists in research fields such as information mining, natural language processing, and genomic informatics.
What Specified? - common and local specifications -
- Shared memory node having 16 cores of 64bit x86 and 32GB memory of 40GB/sec.
- A bundle of inter-node links of 5GB/sec
- Red Hat or SuSE Linux
- Fortran, C and C++ with OpenMP and autoparallelizer
- MPI of 4GB/sec and 8.5 μsec RT latency
- Basic and/or standard benchmarks
- FP performance of components and system
- Number of nodes and total memory capacity
- System-wide network configuration
- Storage system capacity and its performance
- Batch job scheduler and system management
- Commercial libraries and applications
- Site-specific application benchmarks