ATIP99.029 : Climate Modeling and HPC

ASIAN TECHNOLOGY INFORMATION PROGRAM (ATIP) REPORT: ATIP99.029 : Climate Modeling and HPC To: Distribution From: reports@atip.or.jp This is file name "atip99.029" Date: 13 Apr 1999 ATIP99.029 : Climate Modeling and HPC ABSTRACT: This report summarizes the International Workshop on Next Generation Climate Models for Advanced High Performance Computing Facilities, held 1-3 March, 1999 in Honolulu. The report emphasizes the programs, policies, hardware and software systems development occurring in Japan. KEYWORDS: Computer Software/Hardware, Defense Applications, Environmental, Government Policy on Science & Technology, Conferences, High-Performance Computing COUNTRY: Japan

START OF REPORT ATIP99.029

Copyright (c) 1999 by the Asian Technology Information Program (ATIP) This material may not be published, modified or otherwise redistributed in whole or part, in any form, without prior approval by ATIP, which reserves all rights. Climate Modeling and HPC (ATIP/Japan) CONTENTS 1. INTRODUCTION 2. PRESENTATIONS 3. HPC TRENDS 4. EARTH SIMULATOR 4.1 Organization 4.2 Equipment 4.3 Application Software 4.4 Comments 5. VENDOR PRESENTATIONS 6. US PROGRAM 7. CONCLUSION 8. CONTACTS

1. INTRODUCTION The international climate modeling community is composed of a small collection of scientists who are dedicated to understanding changes in climate and weather by the use of sophisticated physical and computer models. The latter include complex physics, chemistry, heat transfer, fluid flow, etc., combined with numerical algorithms to convert physical ideas into computer programs that produce results that can be tested against reality. The goal of such research is to use existing data of current weather and climate to predict short and long-term trends. Such predictions have tremendous economic impact; being able to make accurate forecasts about rain tomorrow, a wet spring, or a late frost, can be directly translated into decisions concerning transportation, farming, clothing, sports, logistics, etc. In addition, the strategic issue of global warming can be studied and, if verified, may be mitigated. In addition to data gathering, the tools of the modeling community include advanced techniques in the physical and mathematical sciences. The need for very powerful computing is also apparent. Weather and climate modeling have always been key applications that are frequently used to measure the capability of new computing systems, and modeling requirements always outstrip existing equipment. The introduction of parallel computing has led both to new challenges and new opportunities for the modeling world. Challenges are due to the various and changing computer architectures that need to be understood and utilized; challenges are also associated with processing, representation, and interpretation of vast amounts of data. Opportunities occur because of increasing computing power per dollar, and by collaborations brought about by new Internet and other networking technologies. A three day workshop, "International Workshop on Next Generation Climate Models for Advanced High Performance Computing Facilities," held at the University of Hawaii's East West Center, 1-3 March 1999, was a good opportunity to examine Japanese and US approaches to some of these issues and especially computing technologies, which are different. This report summarizes this workshop, with a focus on the new computing developments.

2. PRESENTATIONS The workshop was an outgrowth of discussions at the 6th Japan-US Workshop held March 1998, as well as collaborations between Dr Maruice Blackmon at the US National Center for Atmospheric Research (NCAR), and Professor Akimasa Sumi at the University of Tokyo's Center for Climate System Research (CCSR). This is the second year of climate-model-specific meetings between the US and Japan. The workshop goals were to examine the next generation model needs as well as next generation computing requirements and trends. There was a single track of speakers (about 30), primarily from US or Japanese Government organizations, followed by one vendor session, where speakers from IBM, SGI, NEC, Fujitsu, and Hitachi presented current and future product offerings. There was a very small exhibition sponsored by the vendors where participants could ask more detailed questions. Approximately 75 participants attended, primarily from the US and Japan, but also including a few from Europe. The choice of the University of Hawaii's (UH) East-West Center to hold this workshop is symbolic in that it represents a compromise (but attractive) destination for participants from both Japan and the US. In addition, UH has considerable climate modeling work, and in particular, a new International Pacific Research Center (IPRC), that focuses on Asia-Pacific climate models, emphasizing numerical models and diagnostic studies. Of peripheral but significant interest, this year's Japan Prize is given to a member of the University's CS Department, for work on error correcting codes. This topic is relevant, as it relates to large scale data transmission over networks. Workshop Program: http://www.tokyo.rist.or.jp/workshop/program.html

[The remaining sections of this report are available to ATIP subscribers] Tokyo Office: Asian Technology Information Program (ATIP) Harks Roppongi Building 1F 6-15-21 Roppongi, Minato-ku, Tokyo 106 Tel: +81 3 5411-6670; Fax: +81 3 5411-6671 U.S. Office: Asian Technology Information Program (ATIP) PO Box 9678 Albuquerque, New Mexico 87119 Tel: (505) 842-9020; Fax: (505) 766-5166 For further information Send email to : info@atip.or.jp Access WorldWideWeb Site : http://www.atip.or.jp/ ATIP: An autonomous US non-profit, that collaborates with other public, private, and educational organizations, worldwide, to analyze and disseminate information about Asian science and technology.

[Complete ATIP reports on Asian Science and Technology go to subscribers and collaborating organizations by direct distribution, or via electronic access. These contain text and often, charts, graphs and pictures. Reports for unrestricted distribution often contain summarized, or abstracted information. Sponsors can also obtain specific follow up information - including copies of proceedings, selected papers, exhibition particulars, updates, translations, query searches, etc. Contact ATIP at INFO@ATIP.OR.JP]

END OF REPORT ATIP99.029a

[The remaining sections of this report are available to ATIP subscribers] 3. HPC TRENDS From the perspective of most workshop participants, the key discussions were about new computers and associated software that were being developed. (Not everyone would agree. At least one US speaker commented that he "didn't want to know anything about computers, wasn't interested in them, and only wanted to solve climate problems." This didn't seem to be a widely held opinion, however.) Most relevant to ATIP readers, and the focus of our report, were discussions and presentations of Japanese efforts through government and industry to develop advanced computing capabilities. For questions more focused on modeling topics please contact us directly. More powerful computers are required for better modeling. As A.Semtner (Naval Postgraduate School) puts it, "Truly definitive prediction of climate variations and regional climate change can be undertaken as highly parallel machines sustaining 10 to 1,000 teraflops become available." Scientists adapt their research around the tools at their disposal. At the same time they strive to improve and invent new tools. A large fraction of the climate modeling community likes the vector-based shared-memory systems that many of their programs were developed around, and Japanese vendors have emphasized that approach. New microprocessor systems are being utilized more and more, especially in the US, and some impressive results are being shown as programs are being redesigned for these architectures. The longer term economic trend seems to favor technologies that can easily scale and also be leveraged across a broad base of users (customers), including both science and industry groups. To get to the highest levels of performance, hierarchical architectures are being utilized. Although each vendor has a distinct approach one thread was that of clustering. In fact, B.Buzbee (formerly of NCAR) gave a careful opening speech on trends in HPC Architecture, that was subtitled, "Is there a cluster in your future?" He provided estimates of costs based on assumptions of real program efficiency of about 10% for the US ASCI Blue system (3.5GFLOP/M-US$), and 35% efficiency for NEC's SX-5 (4.5GFLOP/M-US$). For the future, he projects the following. Using microprocessor based systems. In 2001: 2500 PEs/TFLOP at a cost of US$71M/TFLOP In 2004: 625 PEs/TFLOP at a cost of US$18M/TFLOP Using vector processing systems. In 2002: 90 PEs/TFLOP at a cost of US$55M/TFLOP In 2005: 23 PEs/TFLOP at a cost of US$14M/TFLOP His conclusions were that, using micros or vectors, "many climate centers will be able to afford sustained TFLOP performance within a few years (2003-2005)." However, Buzbee feels a key unanswered question is how to deal with petabytes of data, and particularly data migration issues, with data rates only increasing by about 4x recently. Milton Halem from NASA (not at the workshop) has also articulated these same concerns in a recent paper for RCI.

4. EARTH SIMULATOR 4.1 Organization The Japanese government program aimed at providing those high performance (e.g., multi TFLOP) capabilities is titled Earth Simulator (ES) or Global Simulator (GS) and nicknamed GS40 internally. GS is actually a three part project to promote global change prediction R&D -- earth observation (generating data), basic research (physics, etc.), and computer simulation. The fundamental goal is large scale computational simulations in earth science. This includes the following. Atmosphere-Ocean phenomena (Global climate change, weather forecasting, etc) using both high resolution global and regional models Solid earth phenomena (Mantle/core convection, plate/seismic tectonics, etc), with obvious application in Japan to earthquake generation processes. This workshop was organized and sponsored by the Research Organization for Information Science and Technology (RIST). This is a new Japanese entity (1995) established by Japan's Science and Technology Agency (STA), which provided the WS financial support. Although not well known in the West, over the past few years, STA has established itself as an important organization, especially in the computing arena with its support in the early 1990's, under Dr Miyoshi, for the Numerical Wind Tunnel (NWT). This computer, installed at Japan's National Aerospace Lab was a prototype that evolved into Fujitsu's VPP series computer. STA's current efforts are similarly focused, this time on what they call "Earth Simulator" (ES), and Miyoshi is involved in this as well. STA is the Ministry-level agency that supports JAERI (Japan Atomic Energy Research Institute), NASDA (National Space Development Agency), and JAMSTEC (Japan Marine Science and Technology Center). JAERI in turn supports a research lab called the Center for Promotion of Computational Science and Engineering (CCSE), that has several small parallel systems installed. GS is mostly a cooperation between JAERI and NASDA, although JAMSTEC has a lesser role. Administratively, RIST is the core organization associated with GS (RIST belongs to STA directly rather than to any of its subsidiary organizations, although staff working there have been detailed from some of the latter). Its primary mission is to develop parallel computing technology, and especially for the GS. A second mission is to develop collaborations with other research organizations in the context of simulating complex phenomena of the earth. RIST is a management organization. The closest analogy in the US (but perhaps not a perfect match) might be the US DOD HPC Modernization Office. Its President, Harumitsu Yoshimura, is a lawyer who spent more than 30 years inside STA. RIST does not seem to actually control the budget for the ES project; instead NASDA and JAERI are more directly involved. In fact, the current RIST Director's office is (a fairly empty) suite of rooms inside the JAERI offices in central Tokyo. A larger and much busier place is CCSE, which is nearby, where actual software development takes place. (CCSE has about 75 people packed into one large room, desks abutting, all pounding away on their workstations (most using flat panel displays to save space); the head of the office sits in one corner. Quite typical for a Japanese workplace.) Yoshimura tells us that the two organizations will be moved together shortly. An interesting question is why NASDA and JAERI are spearheading the ES program, rather than Japan's Meteorological Agency or its Geological Research Institute, as these appear to be more directly connected to the topic (in fact, nobody from either NASDA or JAERI made presentations at the workshop, although there were several representatives from JMA). We suspect the answers are a combination of experience with advanced computing, and political positioning; NASDA and JAERI are larger and better known organizations, and it is likely that the system being developed can be used for other purposes inside them. Another question is the link, if any, between the Earth Simulator project and the Japanese Defense Agency. Yoshimura claims that RIST has nothing to do with the Defense Agency and that the Agency does not do any development work, but only procures systems from vendors. It is obvious that the Earth Simulator is a project the Japanese see as competing head to head with the US ASCI Program, as it will give the Japanese similar computing capability as the US (see below). Although under the current political climate Japan would not use this for military purposes, by giving NEC the capability to develop and build such a machine, the Japanese Defense Agency will have the option of buying one when and if they need to.

4.2 Equipment The Earth Simulator computer represents the third branch of the GS project. This large parallel computer will be the platform for the development of parallel software for predicting global climate change, hopefully with high resolution. STA wants to develop new climate models that allow studies of global warming with less than 30km horizontal resolution allowing parameter studies, for either one-month or longer term climate/weather predictions. Building the hardware, while formidable, is not the only issue. Making very large systems usable is a challenge both at the system software level (compilers, debuggers, etc) as well as programming strategies for taking advantage of the raw power and making effective use of scientists time.) Computer models need to be developed, including more sophisticated physical process modeling, techniques for storing, processing, transferring, visualizing and manipulating terabyte and even petabyte data sets. Concerning software, the primary project, titled, NJR, is a parallel coupled atmosphere-ocean climate model. The ES will be operational in the second quarter of 2002. The target hardware will generate a sustained performance of over 5TFLOPs on an atmospheric circulation program. The architecture is a MIMD distributed memory vector parallel system with 640 nodes. Each node contains 8 vector processors that are tightly connected via 16GB of shared memory. Peak performance will be 8GFLOP per processor, or 40TFLOP for the complete 640 node system. The nodes will be connected by a single stage crossbar with 16GB/sec x 2 data transfer capability, with 3-6microsec latency. A single node can reside in a one-meter square cabinet, produces 64GF, holds 16GB memory and requires 8KVA for power. Technical issues limit the number of nodes that can be connected via a single stage crossbar to somewhat less than 1000, so this design, unless it is layered, does not have ability to scale beyond this configuration. The current design will require 82,000 cables, almost 3,000km of wiring, and weight 217 tons, air cooled. It will occupy a new building outside Yokohama and be placed in a room 50x60meters. Total power requirements for the building will be 15,000KVA, including about 7,000KVA for the hardware. Operators and staff on sight will be in the range 40-50. The GS hardware will cost approximately US$400M, the building and associated support will cost about US$60M. Japanese estimates are that the total project cost will be in the range of US$500M. Using this system, Japanese scientists are hoping to run even larger problems than targeted in the specs. For example, Dr Keiji Tani, who is the Principal Scientist at JAERI feels that to make the type predictions he wishes, will require running global, regional, and local models at 5-10km, 1km, and 100m mesh resolution, using 100-200 vertical layers, and computing at a time step about 1/10th of that used today. He also estimates that using a 10km mesh with 300 words per grid point will require 8TB of data. In actual usage, the system will be organized into 40 clusters of 16 nodes each. One cluster will contain very large disks for user activities, including debugging, etc. The remaining 39 clusters are back-end, batch type, without user disks. NEC is the supplier, and this will certainly be the design of future SX systems (SX-5/6). The company expects to place a complete CPU on one LSI 100mm^2, 500MHz, with 10M transisitors/cm^2, using 0.18 or 0.15mu rules, and with 4,000-5,000 pins per chip. High density packaging, assembling and cooling (heat tubes) will be improved compared to current technology, according to NEC. System software will be NEC's OS which overlays Unix. Additional software will include MPI-2 and HPF-2, with extensions proposed by the Japan Association for HPF. There will also be NEC-produced mathematical libraries (matrix routines, random number generators, FFTs, etc). NEC will also supply operational support, i.e., job and file scheduling. All hardware is to be complete by the first quarter of 2002, system software by the second quarter, and applications software by the end of 2002. JA-HPF is a collection of about 20 compiler specialists from vendors, and a similar number of HPC users that began in 1996. In addition to being a forum for sharing experience, developing benchmarks, translating HPF documents into Japanese, the Association has also introduced original extensions to HPF 2.0, including the following. Reduction kind Asynchronous communication User controllable SHADOW Communication schedule reuse for irregular array accesses. More details on JA-HPF may be found at http://www.tokyo.rist.or.jp/jahpf/

4.3 Application Software The NJR software system, mentioned above, is being designed as a general platform for global change simulation. It will be composed of two independent subsystems, one an atmospheric global change modeling package, and the other a ocean global change modeling package. Within each of these there will also be two independent modules for physics and for dynamics. An atmosphere-ocean coupler will be provided to link these parts. In addition there will be a pre and post processor to provide data input from a climatology database (temperature, precipitation, current, salinity, etc) and to analyze and visualize output data. The intention is to make all the modules plug-in, so it is easy to replace or add other physical processes. Portability will be important so the models can run not only on the ES, but on other vendors' systems, on workstations, etc. This not only includes programming constructs, but also data file formats that are in common use (netCDF, GRIB, etc). NJR designers realize that there are many climate modeling programs available, and they intend to continually verify their approach by validating it against existing models. Current work on NJR involves running against standard models for validation. Performance estimates are provided based upon 16GF node, as follows. Benchmark program Number of nodes Execution time (min) T426L50 160 432 T213L31 80 47 T106L31 40 8 This assumes integration for one year of simulated time, and does not include I/O time. Processing speed is assumed to vary with resolution, based on benchmarks from ECMWF. For these programs there are also estimates for data size, assuming 15 variables, 8 byte input and 2 byte output, as follows. Benchmark program Input Data (MB) Output Data (MB) T426L50 4915 1229 T213L31 786 197 T106L31 197 49 In addition to NJR there are two other software projects that are being developed. One is GeoFEM, a Finite Element Method program for simulation of solid earth phenomena. There are substantial computational difficulties in this type of model associated with structural mechanics, thermal hydraulics, significant nonlinearities, phase changes, as well as multiple time and space scales. In addition data sets are huge and highly distributed requiring parallel processing, multi-level data structures, etc. GeoFEM is a five year project that began in 1997. An early version has been run using a cubic model with a uniform pressure load. A 100M DOF 3D elastic analysis was run on the Hitachi SR2201 at University of Tokyo with 1024 processors. Speed up for the 1K processors was 593 relative to running on a single processor (total elapsed time was 5.9 seconds for the solver). Plans are to build an enhanced version of the package between now and 2001, making it available to the public at the end of this period. The second major software project is SPEEDI, a System for Prediction of Environmental Emergency Dose Information. This is an existing program (1995) that predicts dose distributions resulting from accidental discharge of radioactivites from nuclear plants. Current plans are to parallelize this for the ES.

4.4 Comments At the moment, discussion about Japanese computing capabilities is largely academic within the US organizations represented at this workshop. An effort about two years ago, by NCAR, to obtain a Japanese supercomputer from NEC resulted in an investigation of the company's pricing policies, and ultimately lead to heavy penalties against both NEC and Fujitsu for dumping. Japanese vendors are effectively unable to sell to US public institutions and there is little enthusiasm inside the US science community for contesting that decision. On the other hand, there is evidence that Japanese vector supercomputers are effective at solving certain classes of problems. For example, NCAR's Blackmon reported at this workshop that he ran a global warming program at NCAR, requiring about half a year of computing, and then again in Japan, only requiring about two months, even though the latter included more physics. Blackmon pointed out that better computing allows more ensembles to be run and averaged, more cases, model improvements, and increased horizontal and vertical resolution. The philosophy and background for this project was summed up very succinctly by Akimasa Sumi, the Japanese academic who was the co-organizer of the workshop. Sumi pointed out that at the 1998 meeting there was a consensus of the need for new models and that society demands more accurate climate and weather predictions. Scientists feel that the physical basis for climate simulation is still unclear, e.g., the relation between accuracy of forecasts and specification of computational accuracy, and that right now, scientists are simply trying to simulate climate or predict weather "as accurately as we can." Parameters that need to be improved are the vertical and horizontal grid resolution in numerical models. Sumi feels this will reduce computational diffusion and lead to more accurate computation of forcing, and definitely lead to improvement in modeling of dynamical processes; it may also lead to improvement in modeling of wet processes. (Not everyone agrees that going to higher resolution immediately leads to improvements, however. For example, M.Kanamitsu, US Climate Prediction Center, Nat Center for Environmental Prediction, feels that increasing resolution may exaggerate model weaknesses. Nevertheless Kanamitsu agrees that even with existing models, many important problems cannot be done due to lack of computer resources.) This need to improve models and computation ability has now been linked to the ES program. This was a Japanese domestic program, but is now seeking international collaboration. According to Sumi, there seems no dispute about the need for more powerful computers for climate, weather, materials design, CFD, etc., but Sumi's view is that "Japanese can't trust MPPs" and that future machines will involve moderately parallel computers with vector processors. Although machines of this type are being produced in Japan, Sumi feels that companies need to be pushed, and that without forcing them, the companies will not produce the type machines he thinks are needed. This may be correct -- informal discussions with Japanese vendors suggests that they are losing money on their HPC business.

5. VENDOR PRESENTATIONS Regardless of Sumi's feeling, for the moment at least, all three Japanese computer companies are stating that they will be producing systems on the scale of GS. This was made clear during the afternoon vendor track, that included Fujitsu, Hitachi, and NEC, as well as IBM and SGI. These described their corporate philosophy as well as general ideas about their future directions. (One exception was the SGI speaker, K.Mikami, who gave a highly technical talk about finite element flow analysis, and which seemed strangely out of place.)

IBM G.Wenes discussed IBM's RS6000 program and its plans for the future. Interestingly, he began his talk with the statement that "we are not going to build vector computers," clearly responding to the repeated remarks from other participants (especially Japanese) suggesting this was the best way toward high performance computing. In explaining his comments he presented the core of IBM's philosophy. * Market dynamics will not support substantial development that is unique to very large systems. * Development of very large systems will be accelerated to the extent to which smaller clones of the very large class systems can be used effectively for mainstream applications. * Whatever new technology is required for very large class machines must be applicable to mainstream applications. * Designed well, there is a synergy between very large and mainstream systems -- between the high-end and the high-volume. Certainly IBM's SP series machines have matched these points very well, as attested to by their market success, as well as their reach into the HPC community. For the future Genes pointed to the base technologies that IBM will utilize to obtain higher performance, including 0.18 line width and copper interconnects, as well as peripherals such as its recently announced 1" hard disk drive. Genes tried to emphasize the changing role of the interconnect (switch) in HPC developments, which he feels now becomes the central and key ingredient. IBM: High-Performance Computing: Trends & Directions - Paths for Growth

The Japanese vendors have mostly similar stories to tell regarding current products or those to be released this year -- air-cooled CMOS vector processors with 8GFLOP performance, using line widths of about 0.18 micron. These are tightly coupled into shared memory nodes containing a small number of such processors (up to 16) providing 128GFLOP performance, and connected, when needed, into larger systems. While the largest of these systems reach multi TFLOP levels, Japanese vendors are expecting that they will have significant business delivering small systems. Hitachi's approach is somewhat different and closer to US vendors. Nevertheless, the company has just announced the smallest version of their SR8000 system, one node, 8GFLOP in a single cabinet of floor dimension 50×91cm. NEC, the vendor responsible for the development of the ES, has been successful selling their SX series processors, claiming 220 units delivered. Fujitsu has also been successful. The VPP700 series which has 2.4GFLOP/processor performance is several years old and was designed with 6.5ns clock speeds; current technology is 2-3x faster, thus significant improvements are expected. In fact, Fujitsu has installed their newest system, a 63 node VPP800 (also 8GFLOP/processor) at Kyoto University. The company also states that their existing system at the ECMWF will be upgraded shortly to provide 400GFLOP sustained performance on applications. Hitachi's SR8000 uses 1GFLOP RISC processors, tightly coupled; a single node contains nine such processors. The processors use a unique approach, they call "pseudo vector processing" (PVP), that mimics a very large cache or direct memory access. To couple the processors in the node they have developed system software called Co-operative Micro-Processors in a single Address Space (COMPAS). Hitachi claims that PVP and COMPAS effectively provide automatic vector processing within a processor and intra-node elementwise parallelization performance. Internode communication is via crossbar, 1D up to 8 nodes, 2D up to 64 nodes, and 3D for the 128 node system. One of the Hitachi systems is being installed at the University of Tokyo and another is at the AIST computer center in Tsukuba (along with a large new IBM). Further, the company has announced their intention of building upon this design by releasing a 30TFLOP system around 2002 via chip enhancements and increasing the number of nodes. Hitachi has always had very aggressive pricing/leasing rates, so it will be interesting to see if their approach is successful, given the Japanese interest in vectors. Hitachi's Solution to Teraflops Computing Officially at least, Fujitsu stated what was implied from NEC to a certain extent from Hitachi, that "Vector-based Supercomputing is Alive and Well." The links below provide more detail on Japanese systems and plans. NEC: High Performance Computing in Weather and Climate Modeling with NEC's SX Series http://www.fujitsu.co.jp/hypertext/Products/Info_process/hpc/vpp-e/index.html

6. US PROGRAM This report focuses on Japanese activities. However, for completeness we mention some aspects of US programs. The DOE's ASCI (Advanced Strategic Computing Initiative) is the largest program that relates here, and it does have global climate modeling components. However, most directly overlapping is the ACPI (Accelerated Climate Prediction Initiative) that is currently being considered for funding of US$70M. This is to provide TFLOP computing and infrastructure for the development and applications of comprehensive coupled climate models, and involves participation from the US DOE, NSF, NOAA, and NASA. If funded, this will provide 5TFLOP computing by 2001, and 40TFLOPs by 2005. L.Gates, a workshop speaker from Lawrence Livermore National Lab who summarized ACPI pointed out that a national research program was proposed in response to an urgent national need for knowledge of regional climate change. He also mentioned parenthetically, that the US should not be dependent upon non-US models to specify climate change in the US, and that also the US should be competitive in climate modeling. In addition to high-end computing ACPI also provides support for necessary infrastructure, such as visualization, communication, etc. Gates also pointed out that ultimately perhaps the most important goal of the program is to promote the use of climate change information. Gates explained that the ACPI program has four types of challenges -- scientific, engineering, application, and management, and provided the key points for each of these. International cooperation fits in his management challenge category and was one reason for the presentation at this workshop. Gates also remarked that a detailed implementation strategy has been developed and is available for examination.

7. CONCLUSION This workshop clarified Japanese plans for their climate modeling program including the development of both hardware and software. At US$500M the hardware and facilities investment is somewhat comparable to the US ASCI program. Like ASCI, the emphasis is on high-end computing. But there are many differences too, not only in hardware platform philosophy, but also the application focus. One significant difference is the Japanese emphasis on international collaboration, which was expressed explicitly and in several other ways at the workshop. The two national programs do overlap, however, and changes in ASCI, for example Open ASCI, that provide access to ASCI resources for a wider class of problems and researchers, provide additional overlap with Japan's Earth Simulator effort. The climate communities in the US, Japan, and elsewhere around the world, both compete and collaborate. Most of the researchers are less interested in the platforms they utilize than in moving their projects forward; this is obviously not the case for system vendors. From this workshop it is clear that there is a significant amount of duplication with the US, Japan, EC, and others doing similar research; there is even redundancy that shows up between different national labs. However, models are exceptionally complicated, and it is healthy to utilize a variety of approaches so these can be compared. Further each country/region emphasizes local geometry and so local versions are expected. Nevertheless, there are many opportunities to take advantage of common themes. J.Anderson (GFDL) made these points very clearly, making a plea for more sharing, a consistent modeling framework, and development of tools that can be used more broadly. (He also asked for better documentation, and commented that documenting one model took longer than developing the program.) M.Kanamitsu, at the US Climate Prediction Center, National Center for Environmental Prediction expressed the opinion of many climate modeling scientists when he stated that "I would like to encourage that [an] internationally coordinated seasonal prediction research plan is established around [the] Next Generation Computer under development in Japan." We agree that more cooperation would be helpful. The complexity of the techniques being utilized, not only the physics, mathematics, and algorithms, but the model system integration, data access, analysis, etc., transcend national and laboratory boundaries. These should be developed using the best techniques that are applicable, i.e., computer languages, software engineering, project planning, common data and organizational structures, and so forth, in ways that enable the most cost effective reuse.

END OF REPORT ATIP99.029r

Search ATIP	Return to Home Page