CHAMMP NEWSLETTER, VOL. 1 NUMBER 3 Nov. 1991 Dave Bader, (301)903-4328 FTS 233-4328 e-mail bader@oerv01.er.doe.gov bader@oerv01.er.doe.gov (My preference for all who have INTERNET) or D.BADER on OMNET CHAMMP PROGRAM IMPLEMENTATION As can be expected, things are proceding on course with the program implementation , but progress has been somewhat slower than originally anticipated. The Model Development Effort (MDE) that Bob Malone is leading is off to a good start, even though the plan for the MDE hasn't been finalized or reviewed. We want to include the relationship between CHAMMP and the HPCC in that plan as well as provide details of the use of massively parallel resources by the CHAMMP investigators. CHAMMP is partially funding the development of 1 or more High Performance Computing Research Centers that will be established soon at DOE laboratories. Specifics are still being worked out, so stay tuned for the latest developments. SPEAKING OF MASSIVELY PARALLEL MACHINES-- Bob Malone offered the following description of the new CM-5. Los Alamos will receive the first version of the machine with 1024 nodes and the vector accelerator chips. Bob notes that the best description can be found in an October 28 or 29 Wall Street Journal article. However, since the editor was not willing to retype that particular piece, Bob provided the following items that came over e-mail. From: danny@think.com (Danny Hillis) Newsgroups: comp.arch Date: 1 Nov 91 22:50:33 GMT Organization: Thinking Machines Corporation, Cambridge MA, USA Distribution: world NNTP-Posting-Host: godot.think.com Several people have ask me to post a summary of the Connection Machine model CM-5 5 architecture. Here are some of the aspect of the machine that will be of greatest interest to computer architects: 1. The CM5 has some characteristics of both a SIMD and a MIMD machine. It is primarily designed to run single threaded, data parallel software, but takes advantage of multiple threads of execution to do this more efficiently. Like a MIMD machine, each processor is capable of fetching its own instructions locally, but like a SIMD machine there is a special network for broadcasting blocks of instructions and for keeping the processors closely synchronized. This "control network", in conjunction with the compiler and operating system, provide the user with the illusion of a single threaded machine. The system is upward software compatible with previous Connection Machine at a source language level (Fortran, C*, *Lisp). This single threaded model allows it to efficient execute properly written Fortran 77 or Fortran 90 programs without any parallel extensions. It can also execute MIMD programs written with explicit message passing. 2 The machine uses off-the-shelf SPARC processors as a building block, and adds special chips for communications, memory control, and (optionally) vector processing. The 64-bit vector processors have 128 Mflop/s of peak (multiply-accumulate) performance, 32 Mbytes of memory and a peak memory bandwidth of 1 GByte/s. 3. As currently implemented, the networks, I/O, and memory of the CM-5 are all scalable up to 16K processors. (The architecture is designed to scale up to 256K processors, but with the current packaging, cable length limitations restrict it to 16k.) By scalable I mean that the cross section bandwidth of the network grows in proportion to the number of processors. The entire system is synchronous, even though many of the wires are longer than the clock wavelength. This is accomplished by a system of distributed phase-locked loops. 4. In order to scale up to 16k processors, the machine requires a lot of special hardware for failure detection, isolation, and quarantining. The machine has a separate diagnostic network which uses JTAG protocols to run test vectors on each of the custom chips. When a fault occurs, a defective processing or communications node can be electrically switched out of the system. The machine can continue to run while the failed module is replaced, with the failed components mapped out of the address space. 5. The operating system (UNIX) supports both timesharing and partitioning. The timesharing system swaps "in transit" messages out of the network when switching tasks. The system uses Thinking Machines standard 20 Gbyte disk arrays and can perform striped access across multiple disk arrays. 6. The largest CM5 systems sold so far have 1K processors, and are in the $30M price range. The smallest systems sold have 32 processors, and are in the $1M prices range. 7. The CM5 required about three years to design and build. Over that period the team grew from about 30 people to over 200. Many of the designers will be presenting more details of what we have done at conferences over the next year. -Danny Hillis > ---------------------------------------------------------------------- > 11:54 THINKING MACHINES CORPORATION ANNOUNCES CM-5 SUPERCOMPUTER > > CAMBRIDGE, Mass., Oct. 29 /PRNewswire/ -- Thinking Machines Corporation > today announced its Connection Machine(R) model CM-5(tm) supercomputer, the > first with a complete parallel architecture that scales to a TeraFlops in > performance. At the same time, the company announced important advances in > software scalability that bring supercomputing firmly into open computing for > the first time. Among these advances is a collaboration between Thinking > Machines, IBM and Sun to pursue a common scalable programming standard for > scientific computing. CM-5 system prices start at $1.4 million. > "We are the first to combine a TeraFlops architecture with the production > computing features that bring parallel supercomputing squarely into the > mainstream," said Sheryl Handler, company president. "First customers include > the Army High Performance Computer Research Center, Schlumberger, University > of Wisconsin, University of California at Berkeley, and the Northeast Parallel > Architecture Center at Syracuse." > John Sell, president of the Minnesota Supercomputer Center, which installed > the first CM-5 for the Army High Performance Computer Research Center in > August, underscored the commercial potential. "CM-5 is the first highly > parallel supercomputer that can be seriously considered for commercial > production environments." > Thinking Machines' chief scientist Danny Hillis, emphasized the importance > of scalable architecture. "Scalability completely changes the questions we > ask about performance. Since you can always increase performance by adding > more processors, the question becomes how far can it grow, and does the I/O, > communications, and reliability grow in proportion with the size of the > system. CM-5 is the first supercomputer that is genuinely scalable in this > sense." Each CM-5 node is a 22-Mips RISC microprocessor with four vector > pipes providing a total of 128 MFlops peak speed. All components of the CM-5 > system architecture, including software, I/0, and communications networks, > scale in a balanced way up to systems with 16,000 processing nodes. > Until now, industry growth has been hampered by competing forms of parallel > architecture called "SIMD" and "MIMD." Becuase of CM-5's innovative > architecture, users no longer need to compromise between the programming > simplicity of SIMD and the flexibility of MIMD. The new system provides both, > and includes the synchronization hardware needed for data parallel computing. > Programs running different forms of parallelism may be run simultaneously > under control of the system's full timesharing operating system. Larry Smarr, > director of the National Center for Supercomputing Applications in Illinois, > said, "The most exciting aspect of the CM-5 is its universal architecture > which will provide one-stop shopping for 1990's applications developers. I > really admire the long-range vision of Thinking Machines. With the CM-5 we > have a scalable architecture that will carry us all the way to a TFlops." > Hillis expanded on the importance of scalable software. "With today's > networks, scientists use mainframes one minute, workstations the next, and > supercomputers the next. In the past, these performance levels have been > locked into their own proprietary language standards. Open, scalable software > gives users the ability to run the same Fortran on a Sun workstation, an IBM > vector supercomputer, and a TeraFlops Connection Machine system. This means > that third-party software developers will be able to maintain a single source > for all types of machines." DARPA (the Defense Advanced Research Projects > Agency) funded the original development of scalable architecture as part of > the U.S. High Performance Computing Initiative, and is taking a leadership > role in guiding the industry transition to scalable software. > The transition to scalable software is the focus of Thinking Machines' > initiative in common scalable programming standards for scientific computing. > Joining Thinking Machines in this effort are IBM and Sun Microsystems. This > announcement is significant because the three companies span the range of > scientific computing, from workstations to shared memory multiprocessors to > Grand Challenge supercomputers. Thinking Machines Corporation is the > world's leading manufacturer of highly parallel supercomputers and a pioneer > in scalable computing techniques. The company, which is privately held, is > headquartered in Cambridge, Mass., with offices worldwide. > ---- > NOTE: Connection Machine is a registered trademark of Thinking Machines > Corporation. CM-5 is a trademark of Thinking Machines Corporation. > /CONTACT: Tim Browne of Thinking Machines Corporation, 617-234-5525/ > 11:39 EST > > PR NEWSWIRE STORY 28 PAGE 1 OF 3 > > 11:54 NEW INITIATIVE IN OPEN, SCALABLE, FORTRAN PROGRAMMING ANNOUNCED > > CAMBRIDGE, Mass., Oct. 29 /PRNewswire/ -- Thinking Machines Corporation > announced today an initiative to create a common scalable programming standard > for scientific computing. Joining Thinking Machines in this effort are IBM > and Sun Microsystems. This announcement is significant because the three > companies span the range of scientific computing, from workstations to shared > memory multiprocessors to Grand Challenge supercomputers. > Danny Hillis, chief scientist of Thinking Machines, explained the importance > of scalable software. "With today's networks, scientists use mainframes one > minute, workstations the next, and supercomputers the next. In the past, > these performance levels have been locked into their own proprietary language > standards. The future belongs to scalable applications that run unchanged up > and down performance levels." > Dr. David S. Wehrly, director of high performance supercomputing systems at > IBM, positioned his company's supercomputing directions. "Data parallel (such > as the Fortran 90 array language) is IBM's direction toward exploiting > supercomputer parallel accelerators. IBM intends to participate in external > activities focused on languages and environments to exploit parallel > processing." > Thinking Machines and IBM have previously announced a joint development > agreement between the companies. Thinking Machines pioneered data parallel > programming. Fortran 90 array language is widely used on the company's > Connection Machine(R) supercomputers. > Jon Kannegaard, vice-president of SunPro, a Sun Microsystems Inc. business, > talked about the importance of open scalable software to his company. "Sun > makes workstations and offers shared memory multiprocessors. We are enthused > about Thinking Machines-compatible array extensions in Fortran 90 as a > standard, scalable, programming model for these multiprocessors. It is the > right way to program shared memory machines as well as massively parallel > machines." > Scalable software techniques complete the open computing revolution that > started with portability. Portability allows the same program to run > unchanged across vendors. Scalability allows the same program to run > unchanged up and down performance classes. > Thinking Machines Corporation is the world's leading manufacturer of highly > parallel supercomputers and the pioneer in scalable computing techniques. The > architecture of its CM-5(tm) supercomputer is the first to scale to a trillion > operations per second (TeraFlops), the performance level required by the new > set of computing applications called the Grand Challenges. > ---- > NOTE: Connection Machine is a registered trademark of Thinking Machines > Corporation. CM-5 is a trademark of Thinking Machines Corporation. > /CONTACT: Tim Browne of Thinking Machines Corporation, 617-234-5525/ > 11:38 EST > > CHAMMP PILOT PROJECT VOLUME The compilation of pilot project final reports is being edited and will, with any luck, be available in January. Only 1 month late, so things are getting better!!! NEW SCIENCE TEAM MEMBERS The following investigators received the first round of DOE laboratory science team projects resulting from the proposal solicitaion/review process last spring. The list of first round external grant recipients appeared in the last newsletter. Second round external science team members will selected in December or January. J. Baumgardner, R. Smith and J. Dukowicz (LANL) Comaprison of Methods for Solving the Spherical Shallow-Water Equations on the Connection Machine" J. Drake (ORNL), D. Williamson (NCAR) and I. Foster (ANL) "Improvements to Parallel Algorithms for Climate Modeling" D. Denbo and E. Skyllingstad (PNL) "Improved Parameterizations for Deep Convection Processes in Ocean General Circulation Models" I. Foster (ANL) and D. Williamson (NCAR) "Computational Aspects of Climate Model Physics" L. Margolin and D. Holm (LANL) "Assessing Predictability in Numerical Modeling of Ocean Dynamics" R. Smith and J. Dukowicz (LANL) "Massively Parallel Global Ocean Modeling" H. B. Stewart, P. Micheal and R. Peierls (BNL) "Dynamical Structure of Climate Systems" D. Wuebbles and J. Penner (LLNL) "Atmospheric Chemistry and Climate Predictability: Towards and Advanced Climate Model" FIRST SCIENCE TEAM MEETING Mike MacCracken is currently finalizing an agenda and suggested dates for the first science team meeting. Science project PI's should be hearing from Mike in the next several weeks. JASON Meeting JASON held a briefing for the CEES agencies on Nov 4 in Washington, DC on their summer session on climate change. Steve Koonin of JASON reported the results of the JASON review of CHAMMP that was conducted last summer. The conclusions were generally favorable, but JASON is still not convinced that climate prediction with GCM based modeling is achievable for decade to century scale climate change. JASON committee members are still finalizing their written report and we have not yet seen it. MEETINGS Atlanta, GA January 6-8, 1992 American Meteorological Society's Symposium on Global Climate Change held in conjunction with the AMS Annual Meeting and several other conferences. Talks will be given by Ari Patrinos on ARM and Dave Bader on CHAMMP, as well as several other new programs in global climate change during a session on January 6. CONTRIBUTIONS STILL ENCOURAGED Anyone who wants to contribute to the newsletter is encouraged to do so. CHAMMP Contacts Dave Bader, CHAMMP Program Director (note corrected e-mail address) (301)903-4328, FTS 233-4328 (NOTE NEW COMMERCIAL PREFIX) bader@oerv01.er.doe.gov Mike MacCracken, CHAMMP Chief Scientist (510)422-1826, FTS 532-1826 (NOTE NEW AREA CODE) mmaccracken@llnl.gov Bob Malone, CHAMMP Director of Model Development (505)667-5925, FTS 843-5925 (Bob's Number will change soon as well) rcm@lanl.gov