Feeds

Fujitsu's King K thrashes Top 500 rivals

Supercomputer's amazing 10-plus petaflops with 706,000 cores

Build a Business Case: Developing Custom Apps

Fujitsu's K computer has confirmed its place as emperor of the Top 500 supercomputer list with an incredible 10.51 petaflops.

The K computer – RIKEN, a Japanese government science and technology research institute, after the Japanese word for 1016, "Kei" – was conceived in 2006. Detailed design took place from 2007 to 2009, when manufacturing started, with the system being delivered in stages. Performance tuning started this year, with an 8.126 Pflop rating in June. Full service is scheduled to start in November.

The petaflop number has just increased by 29 per cent, which solidly cements the K computer's place at the top of the SuperComputer Top500 list. The second system is a 2.566 Pflop Tianhe-1A supercomputer at the Chinese National Supercomputing Centre in Tianjin, made by NUDT. Third is a Cray Jaguar at the US Oak Ridge National Laboratory, rated at 1.759 PfLops.

The numbers involved are bizarrely large. The K computer has 11PB of local file system storage and over 30PB of global file system storage, mostly Fujitsu Eternus arrays. There are 864 racks, housing 88,128 SPARC64 VIIIfx processors: each one a node in the system. Each processor has 8 cores, running at 128Gflops and 2GHz, connected by a Tofu 6D mesh torus interconnect that can support more than 100,000 nodes. That means a grand total of 705,024 cores. This baby is a monster.

It is installed at the RIKEN AICS (Advanced Institute for Computational Science) multi-storey site in Kobe, Japan and the project has been sponsored by Japan's Ministry of Education, Culture, Sports, Science and Technology. It seems obvious that, at this scale, no one company can privately develop such a huge processing engine and government intervention and funding help will be essential to go past the 10.51 Pflop mark.

It seems equally obvious that no private organisation could afford to buy an 88,128 node K computer. These colossally expensive machines are going to be national or international resources for hire.

El Reg was given a briefing by Masahiko Yamada, the President of Fujitsu’s Technical Computing Solutions Unit. He said that the system would be used for research in five areas:

  • Life sciences and drug manufacture;
  • New material and energy creation;
  • Global change prediction for disaster prevention and reduction when earthquakes, hurricanes and tsunamis occur;
  • Industrial innovation such as the complete design and simulated testing of cars and airplanes; and
  • Deep physics, such as the origin of matter and the exploration of the universe.

Saving Japan

Last year's offshore earthquake and tsunami that devastated the coastline north of Tokyo has put global change prediction into particular focus. The idea is that the K Computer could be used to model the effects of tidal waves hitting towns, villages, roads, railways and industrial installations on the coast and see how protecting them with various measures could help them resist tidal waves and recover from their effects faster.

This would help minimise the loss of life and the tremendous financial costs involved.

A life sciences project involves checking how different drug design molecules could fit into so-called pockets on the surface of bacteria and deal with them effectively. Yamada said there is ongoing joint research in this field with a Professor Hideaki Fujitani of Tokyo University.

The Linpack result involved testing which ran for 30 hours. Yamada said that other supercomputer systems would only run for a maximum of around two hours and then stop, because errors somewhere in the system caused the job to stop. With thousands upon thousands of cores and interconnect points, the likelihood of an error is quite high and mainframe-class design reliability is needed which, of course, this Fujitsu baby has.

Yamada said that the K computer was highly efficient, with a 93.17 per cent Linkpack efficiency rating.

Asked about flash, he said that it was not present in the system. Back in 2006 when it was conceived flash storage did not have anything like the predominance it is now attaining.

Now that the system is nearing operational readiness it needs users.

Yamada said the Japanese government is promoting joint research efforts between Japanese universities and institutions to develop applications of high-performance computing in new fields beyond the traditional ones.

It is necessary to extend the research efforts to solve issues in different countries with a global collaborative scheme. Fujitsu is pushing this idea energetically.

Having spent what's likely to be over a billion dollars of its own and the Japanese government's money it is understandably keen that the almost three-quarters of a million cores don't sit there running Grand Theft Auto because there aren't enough customers. ®

Using blade systems to cut costs and sharpen efficiencies

More from The Register

next story
Amazon sues former employee who took Google cloud job
Alleges breach of non-compete clause in contract
New research: Flash is DEAD. Yet resistance isn't futile - it's key
Electro-boffin may have SAVED the storage WORLD
THE GERMANS ARE CLOUDING: New AWS cloud region spotted
eu-central-1.amazonaws.com, aka, your new Amazon Frankfurt bitbarn
Aereo presses pause on 'tiny antenna' TV-streaming service
Vows to fight on after US Supreme Court ruling
IBM promises mainframes on tap as SoftLayer lands on London
Unbelievable... Chessington's nowhere near London
Dimension Data cloud goes TITSUP down under... after EMC storage fail
Replacement hardware needed as Australian cloud flops for 48-plus hours
WD adds Miracast support to feed-slinging tellybox gizmo
Small glitch: You still can't watch stuff from iTunes on it
IDC busts out new converged systems charts, crowns Oracle as Platform King
Nutanix/Simplivity not shown - but they're there. Oh yes
Cloudera, MapR, IBM, and Intel bet on Spark as the new heart of Hadoop
Screeching sound heard amid Spark stampede as Cloudera changes opinion on Hive
prev story

Whitepapers

How modern custom applications can spur business growth.
In this whitepaper learn how to create, deploy and manage custom applications without consuming or expanding the need for scarce, expensive IT resources.
The Power of One eBook: Top reasons to choose HP BladeSystem
Only the Power of One delivers leading infrastructure convergence, availability and scalability with federation, and agility through data center automation.
The Essential Guide to IT Transformation
ServiceNow discusses three IT transformations that can help CIO's automate IT services to transform IT and the enterprise.
Maximizing your infrastructure through virtualization
Virtualization continues to be one of the most effective ways to consolidate, reduce cost, and make data centers more efficient.
Build a Business Case: Developing Custom Apps
In this whitepaper learn how to maximize the value of custom applications by accelerating and simplifying their development.