The "business as usual" model will inevitably increase global warming and consequently destruction, faster than previously predicted. The European Union (a high polluter) has taken onboard the IPCC conclusions and the findings of Nicholas Stern, an economist and author of a forceful report commissioned by the UK government, on the cost of Global warming. The EU already agreed to drastically reduce green house gas emissions by 20% between now and 2020.

The recent Asian Pacific Economic Community (APEC), which includes the USA, China, India, Australia, the biggest polluters are however dragging their feet. This is likely to reduce the pressure on EU countries to deliver their targets. In the meantime the devastation of property, infrastructure and threat to life continues unabated. Heat waves, forest fires, drought, flooding and hurricanes are becoming everyday news items.

It was in this climate that on September 9 to 13, 2007, about eighty meteorologists and HPC experts from large scale computing centres from eleven countries attended the bi-annual CAS2007 workshop on the use of HPC in meteorology, held at the idyllic Imperial Palace Hotel, Annecy, France. The workshop was organised by the National Centre for Atmospheric Research (NCAR), USA. This excellent small and friendly workshop provided a tour de force in meteorological and computing techniques by active practitioners, some of them IPCC contributors, striving to maximise the latest HPC technology to refine and improve their climate prediction models.

Refinement of models is an urgent requirement for the development of realistic mitigation strategies to address the potential catastrophic consequences of global warming. These talks were augmented by speakers from broader scientific centres of excellence like CERN, NERSC and ORNL and from funding bodies such as the NSF, in the USA.

Most presenters came from sites in the USA with large Cray XT3/4, IBM Power5/6 and Blue Gene/L systems, while the European and Australian contingent included a strong representation from sites with large-scale parallel vector NEC SX-8 systems. The main vendors in the HPC space described the next products, their vision for Petaflop/s and the technology advances needed for Exaflop/s systems. What became abundantly clear during this workshop is that "the business as usual model" be it in human activities as a whole or in developing computing technologies, is not a realistic option and radical new approaches are needed. This article highlights a few of the many climate and technology issues raised by presentations given at this workshop.

There were 38 presentations in 3.5 days, some describing the Grid's enabling potential for international collaboration, such as CERN and also within the community of climate system modelling (CCSM). The talks were crammed with technical information on how to use parallel supercomputers for computation using mathematical models, which describe climate/weather patterns over time. These were interspersed with weather maps and video pictures from simulations and compared with satellite pictures of actual weather events.

Why are meteorologists doing all this Earth System Modelling and what is the urgency? As stated above, dramatic flooding and other climate change extreme events are happening and frequently reported in the press and television. For example, it has just been reported that satellite images show that the North West passage connecting the Atlantic and the Pacific oceans is free of ice making it navigable for the first time since records began. Also seventeen central African countries are currently flooded with millions of peoples' homes and crops devastated. "It is common knowledge that it is the countries of the south that stand to be hardest hit by global warming when at present it is the countries of the north that are the biggest polluters", stated Nicholas Stern.

Climate simulations show that green house gases attributed to human activities is causing an increase in the Earth's average temperature. Consequently, fires from intensely hot summers and flooding from heavy rainfall are becoming more common. These images of devastation and the economic aftermath are injecting a political dimension into the proceedings.

A number of talks dealt with prediction and mitigation strategies, to deal with devastating events such as flooding and hurricanes. The costs of these events are enormous. Hurricane Mitch caused the death of over 9000 people in Nicaragua mostly from flooding and landslip.

Greg Holland from NCAR gave a stimulating keynote talk titled: "Anthropogenic influences on intense hurricanes". This presentation was rigorous not only on observed hurricanes, but also on the scientific evidence for causal attribution.

He went on to say that apart from direct impact, indirect impacts arise from forecast uncertainty, design and preparations and imperfect knowledge of cyclone parameters. Coastal impacts from tropical cyclones include harbour damage, houses and crops destruction, forest damage from wind, waves, flooding and landslips. Excluding the loss of life, the direct damage from hurricane Katrina was about $80 billion dollars and an additional $40 billion as indirect costs to the USA. About 95% of the oil and gas production in the Gulf was disrupted and about 150 oilrigs were lost. Tornadoes and flooding reached as far away as Quebec and people were displaced from most coastal states. Government recovery costs were $10-15 billion. The CPI impact was around 1.4% to 2.3%. The cost per household ranged from $140 to $230. The total CPI cost was estimated to be: $16 to 26 billion. Reduction in economic growth rate was about 1%, but this was compensated by subsequent overshoot in the economy. In the USA, 50% of the population live within 50 miles from the coast and the cost for evacuating one mile of coast is about $1 million per day. Note that recovery from a storm impact takes years.

Greg went on to explain that the IPCC position on hurricanes: "It is likely that increases have occurred in some regions since 1970; it is more likely than not a human contribution to the observed trend; it is likely that there will be future increasing trends in tropical cyclone intensity and heavy precipitation associated with ongoing increases of tropical SSTs, is already out of date. He demonstrated with detailed graphs that the bulk of the warming since 1970 is due to anthropogenic effects. He reviewed Atlantic SST and atmospheric modes of variation and demonstrated that these are not accounted by natural variability. His conclusion was unequivocal: "Anthropogenic climate change is substantially influencing the characteristics of North Atlantic tropical cyclones through complex ocean-atmosphere connections and may be influencing other regions".

The above view was re-enforced by Dr. Warren Washington, in his talk titled: "Computer modelling of the 21st century and beyond". From the 1970s, Warren sat on a committee, which advised six US presidents on climate issues. He is also heavily involved at NCAR in "The Community Climate System Model" (CCSM), which has produced one of the largest data sets for the IPCC fourth assessment. He repeated that: "As a result of this and other assessments, most of the climate research science community now believes that humankind is changing the Earth's system and that global warming is taking place". He also told me that every question raised by sceptics was thoroughly reviewed and rigorously refuted. "Natural events do not explain global warming. The smoking gun is human emissions and once included the warming can be reproduced from year to year".

According to Paul Crutzen, another IPCC participant: "The only criticism that could be made of the IPCC report is of it being too cautious".

For HPC vendors the good news is that there is a lot of new work to be done requiring oodles of computing power. The prediction needs include global responses to external forcing (anthropogenic, solar, volcanic ...), internal oscillation of the climate system, internal tropical modes (ENSO, intra-seasonal and waves), external to tropics modes (Atlantic/Pacific multi-decadal modes) and so on.

The computing requirements are for data assimilation, modelling internal oscillations, prediction of external forces and hurricane/climate feedback. A system delivering 200Teraflop/s of sustained performance would be appreciated and usefully utilised as of today.

In fact most sites pursuing climate change research are well endowed with computer resources. Three years ago they were lucky to muster 0.5Teraflop/s of sustained performance. Current procurements in progress have minimum requirements of around 10Terflop/s of sustained performance in phase one followed by doubling or more by 2009. The climate applications of course can use an order of magnitude more power now without waiting for the end of the decade. The tantalising fact is at least one system capable of delivering 200Teraflop/s of sustained performance with less than 9,000 processors will be available from Japan early next year, but it is unlikely to be sold in the USA.

In the next few years, the CCSM will be further expanded to include reactive troposphere chemistry, detailed aerosol physics and microphysics, comprehensive biogeochemistry, and ecosystem dynamics, and the effects of urbanization and land use change. These new capabilities will considerably expand the scope of earth system science that can be studied with CCSM and other climate models of similar complexity. Higher resolution is especially important near mountains, river flow, and coastlines. Full hydrological coupling including ice sheet is important for sea level changes. It will include better vegetation and land surface treatments with ecological interactions as well as carbon and other biogeochemical cycles.

For example, Dave Randal has a five years NSF grant to study clouds using high-resolution models. Clouds are central to Earth sciences, climate change, weather prediction, the water cycle, global chemical cycles and the biosphere. Dave stated: "We are being held back in all of these areas by an inability to simulate the global distribution of clouds and their effects on the Earth system". The need for high resolution catapults this application into needing Petaflop/s computing performance.

The computer requirements, for the next generation of comprehensive climate models, can only be satisfied by major advances in computer hardware, software, and storage. The classic climate model problems with supercomputer systems are: The computers (with the exception of vector systems) are not balanced between processor speed, memory bandwidth and communication bandwidth between processors, including global computational needs. They are more difficult to program and optimize; it is hard to get I/O out of computers efficiently and computer facilities need to expand archival data capability into the Petabyte range. There is a weak relationship between peak performance and performance on actual working climate model programmes.

Thus with sustained performance Teraflop/s computing now on stream, meteorologists are moving from climate to Earth System Modelling (ESM). This is because feedback loops of climate system with other relevant systems like ecology and socio-economy are not negligible. Climate modelling is not possible without proper representation of these systems. Earth System Modelling is: Multi scale (time and space), multi process, multi topical (physics, chemistry, biology, geology, economy ...). It is both very compute and data intensive. Some people claim it requires several orders of magnitude more computing power to tackle the problem. Petaflop/s and Exaflop/s are therefore eagerly awaited.

Stefan Heinzel, director of Garching computing centre of the Max Plank Society (RZG), Germany gave a keynote titled: "Towards Petascale computing in Europe - A challenge for applications and hardware vendors�. He listed the present Petaflop/s projects and their likely hardware architectures. The Riken project in Japan, the Cray Cascade and the IBM PERCS in the USA were mentioned. Doing the math he indicated that one Petaflop/s of sustained performance would need hundreds of thousands of cores. Even for vector machines, although somewhat smaller numbers, it could be around 50 thousand of cores. The question is how does one deal with O(50000) parallelism using local memory and MPI? What of the CPU memory gap, which keeps widening as CPUs grow much faster per year, than the 7% increase of memory speed.

Transistors on an ASIC are still doubling every 18 months at constant cost, but in the last two years AMD/Intel did not announce significantly faster cores. Performance improvements are now coming from an increase in cores on a processor. Present 4 and soon 8 cores will become standard. Intel already announced an 80-core processor technology. IBM doubled the performance from the Power5 to Power6 reaching 4-5GHz, but using 2 cores. The Power6+ could increase the frequency incrementally, but doubling the frequency of the Power7 is going to be difficult. After year 2015 one envisages about 512-cores (or even more nano-cores).

Sequential applications are of O(1). There is no substantial performance increase delivered by faster cores. It is the same on the desktop and on HPC systems. The snag is that memory speed increases only 7% per year, with no improvement of latency of the cache architecture or memory bandwidth. Current HPC applications can use O(K) MPI tasks, mapped to threads. "Classical" scaling can achieve not higher than O(3,000); while the Blue Gene/L can achieve O(30,000). Higher scalability in the range of O(M) requires new technologies in the processor and in the nodes. An SMP programming model between hundreds of cores requires hardware support for lock mechanisms, transactional memory for atomic updates, new micro architecture for latency hiding and pre-fetch hints, i.e. with "assist threads". The memory bandwidth wall is the limiting factor for scalability of the multi-core architecture.

The File and I/O system needs to support hundreds of thousands of clients. To solve the scalability problem a low latency communication is required. One expects a huge number of files in a single file system. Trillions of files with Terabyte/s transfer rates. For robustness different techniques are also needed, since RAID6 is not an adequate I/O connection mechanism.

For Petaflop/s systems, the Operating System (OS) and middleware need to have awareness of massive parallelism. OS failures have to be reduced. OS jitter impacts can have dramatic performance degradations for applications. Lightweight operating systems or special reduced standard kernels should be considered. There is necessity for interrupt synchronization and dynamic management of various page sizes. Applications should adapt their page sizes dynamically, and a reduction of boot time and time to load an application. Hierarchical concepts should be implemented to solve scalability issues; hierarchical daemon structures are required for supporting hundreds of thousands of clients and by queuing and monitoring systems.

And of course one hits the processor power wall. Smaller cores help, providing more operations per watt. A lower frequency also helps to increase the number of cores, i.e. more operations per watt. This implies higher scalability, but the ratio of sustained performance is an open question. Memory and huge caches use significant power too and so does interconnect. The challenge is to optimize the power consumption of each component.

In summary, the use of multi-core and later nano-core architectures implies a challenge for petascale applications. There is a need to hide the memory wall. Cache architecture and memory have to improve latencies and bandwidth. New synchronization mechanisms have to be realized with SMP parallelism becoming more important. Helper threads will support pre-fetch techniques, but the applications have to improve latency tolerance. The interconnect limit implies new programming techniques are needed. Increase of power consumption is the most critical problem due to budget limitations. The necessity of reinventing parallel computing is an enormous challenge caused by the massive increase of cores in future architectures.

The Cray XT series and IBM Power and Blue Gene systems, as well as other vendors in the USA and vendors such as NEC and Fujitsu from Japan are developing Petaflop/s systems, but how effectively these systems can be utilised is an open questions and fraud with a myriad of challenges.

During the last CAS workshop in 2005 a strong emphasis was placed on data management and the challenges this entails. This time the emphasis was more on power used by supercomputers (carbon footprint) and the facility footprint (space) requirements, which are of most concern.

In my view, global warming is the most pressing challenge of the 21st century and we all need to reduce our carbon footprint and become carbon neutral. Our political leaders should be judged whether "fit for purpose" and held to account on what mitigation policies they enacted to slow down global warming. When Mr. Alan Greenspan, the former federal reserves (US central bank) chairman says in his newly published autobiography that: "the Iraq war is largely about oil", he should have added that it is also an irresponsible way of avoiding taking the economic decisions needed to start mitigating "an inconvenient truth", global warming.

The workshop presentations are expected to be available on the NCAR Website:

http://www.cisl.ucar.edu/dir/CAS2K7/