Weapons Experts Present Experiences of ASCI Programme

On the first morning Paul Messina, the man responsible for managing the Accelerated Strategic Computing Initiative (ASCI), for the US Department of Energy, and the link in the National Security Agency (NSA) on Nuclear matters, described the motivation for the ASCI programme. In a nutshell President Clinton and the US congress have asked the Federal Labs to come up with a proposal for using computer simulations capable of developing and maintaining nuclear weapons without the need to do physical testing.

The programme would also have to encompass maintenance and validation of current stockpiles. In practice the nuclear weapons programme involves more than just making bombs, it involves physical material studies, safety, safe decommissioning and long term safe storage without degradation. It turns out that the DOE spends about a third on integrated computer systems, a third on defence applications, (designs of weapons) and a fifth on design of materials. For these activities they calculated that one needs a 100 Tflop/s to simulate 3-D Transport or 3-D material models. Hence the ASCI requirement was born. In the age of Global instant news, one thing politicians are not prepared to be associated let alone fund, is anything with a whiff of flops in its name, such as TeraFlop/s, hence the measure had to be renamed as TeraOps.

Moore's Law in action to deliver ASCI

The task was how to get from a capability of around 50 Gflop/s sustained performance in the early nineties, to 100 Tflop/s sustained by the year 2005. A simple calculation shows that computer technology, both hardware and software, had to deliver at a rate doubling performance every eighteen months as postulated in the so called Moore's Law. This rate of growth can not continue for ever, as the physical limit when using silicon, is reached at about a tenth of a micron etching distances. The race was therefore on, not only to produce fast enough processors and memories, fast interconnect buses and communications, but also fast storage and the software necessary for the integration of all these components into a coherent reliable environment to deliver the ASCI mission. To put it mildly, this was not just ambitious leading-edge, or cutting-edge but as some people said: it was at the bleeding-edge.

ASCI platforms - World's Most Powerful

Of the three Federal Labs, Los Alamos and Lawrence Livermore design and make weapons where Sandia is the engineering hub validating integrity of ageing weapons and so on. For historical reasons and with their experience of using parallel systems like the 1000 processors NCube-2 machine in the late eighties, Sandia was chosen as the first test-bed for the ASCI programme. They purchased an Intel system, code named ASCI Red, based on 9360 processors, having 1.2 Tera Byte memory and 12.5 Tera Byte storage. This amounts to 3.15 TeraOps peak with the expectation to obtain 1 Teraflop/s sustained.

Los Alamos wanted a cluster of SMPs with a 3 Teraflop/s peak, running UNIX, a high speed interconnect and the ability to switch resources between unclassified and classified work. They opted for an SGI solution code-named Blue Mountain, with single system image over 4096 processors. Soon after the contract was signed things started to go wrong. SGI was unable to develop the fast processor promised and after several years of painful trials they ended up with 6144 processors , 1.5 Tera Byte memory, 76 Tera Byte of storage and with just over 3 Teraflop/s peak.

Lawrence Livermore chose an IBM solution, code-named Blue Pacific, based on the RS6000 SP processors. IBM provided 5856 processors, 2.6 Tera Byte memory, 52.5 Tera Byte storage, delivering 3.89 Teraflop/s peak.

There are plans to upgrade this as part of the next phase (ASCI White) to use 8192 processors, 4 Tera Byte memory, 150 Tera Byte of storage, delivering 10.2 Teraflop/s by summer 2000. Beyond this, the forward path envisages a system with 30 Teraflop/s in 2001 and 100 Teraflop/s in the year 2004-5.

These machines cost about 120 million dollars each, ignoring facility costs. They run the 3-D primary dynamic alpha code at 1 Teraflop/s sustained, typically taking 20 days to get a solution. They have a foot-print of at least 1,200 square metres and use 6.2Mwatts for cooling and power. It was calculated that the electrical energy needed to simulate is of the same order as doing the actual nuclear test.

Diversity Withstands Destructive Power of ASCI

The ASCI programme was on one level a disguised subsidy for US HPC vendors to stimulate R&D, but on the other hand its exclusive focus on weapon developments has distorted the market. It encouraged specialised solutions, such as scaleable computing, and in the USA at least strangled the Parallel Vector Processors (PVPs), based on shared memory.

Of course business necessity has a habit of re-asserting itself, and since PVPs are the only present computers systems capable to solve large sparse linear algebra problems, the scientific and technical engineering businesses, Aerospace and Car manufacturers in Europe, have migrated to the Japanese vendors, such as NEC. Judging from the excellent presentation by Christian Lantwin on the performance of the NEC SX-5, which and I quote: "can deliver transparently and with ease very high performance and has the capability to deliver up to Teraflop/s, in a shared memory environment on mature software", one can see why. The Japanese also have a futures path with their Earth's Simulator project, of reaching 40 Teraflop/s. The case why sparse linear algebra problems are best suited to PVPs was convincingly presented by Burton Smith, the Maverick designer of the Denelcor HEP, the Tera MTA and as of the 4th April the owner of Cray Inc.

Moore's Law runs out of steam by 2010?

It was left to Professor Tony Hey to entertainingly wrap up the conference, and at the same time point out that the future of computing lies with the world wide web and e-commerce, rather than HPC, and Moore's Law will run out of steam on silicon by 2010. The world has to explore and prepare to use new material. He already dubbed the 21st century as the age of nano-technology.

The crowded programme during these two days delivered a lot of state-of-the-art information and probed the limits of today's computer technologies.