Australia's Bureau of Meteorology is looking to replace its high-performance computational (HPC) system next year to improve its weather modelling capabilities.
The agency has approached industry for proposals on how to replace its current Oracle/Sun-based HPC system.
That system has 576 compute nodes rated at around 49 Teraflops using 2.93 GHz Intel Nehalem processors and 24GB memory compute nodes, QDR infiniband interconnect, and Lustre file system, the BoM said in tender documents.
The Bureau has based the specifications for its proposed new system on the Australian Government’s Fujitsu-based ‘Raijin’ supercomputer in Canberra.
Raijin has 57,472 Intel 2.6Ghz Xeon Sandy Bridge cores in 3596 compute nodes, around 160TB main memory and a peak performance of 1.2 petaflops. It cost $50 million to build and $12 million a year to run.
The Bureau is bidding for a supercomputer with 50 percent more capability than Raijin.
The agency collects around 1TB of data every day, and expects this to grow by 30 percent every 18 months to two years.
“HPC systems play a critical role within the infrastructure of the Bureau in support of time-sensitive, 24x7 weather prediction services, severe weather warning services, and emergency environmental responses to events such as floods, tsunami, bushfire, nuclear, volcanic ash, chemical spills, air quality and dust,” the bureau noted in tender documents.
“The Bureau will need to replace its existing HPC system in 2016 to meet the requirements of the current and future meteorological, oceanographic, hydrological and environmental services through deterministic and probabilistic numerical prediction and analysis systems from weather to climate time-scales. The Bureau’s computational and storage infrastructure is critical to the delivery of these services.”
The Bureau approached the market for proposals but will only choose a supplier should it go to tender next year.
Two options were put forward for a replacement — either a single supercomputer with 600 TFlops or more in a “well-balanced interconnect and storage I/O subsystem” located in one off-site data centre; or two identical supercomputers each with 1.2PFlops or more also located in off-site data centres.
The eventual successful supplier would install, maintain and support the new system from next year until 2021, with a 2018 upgrade scheduled to “meet the computational capability and capacity needs of future numerical prediction systems and meteorological services, including the option to replace multi-core processor nodes with many-core processor nodes for improved energy-efficiency.”