Contents of
October Issue
|
Contents of October Issue |
Third GMD - NEC Workshop on Scientific Parallel Computing
Munich 16-10-1996 The third GMD - NEC Workshop on Scientific Parallel Computing in St. Augustin on September 6-7, concentrated, as Professor Ulrich Trottenberg, GMD, explained, on new topics in computational fluid dynamics and structural mechanics. The meeting brought these different engineering fields together. It is a question of the future as to whether we can integrate two application areas which have already demonstrated the benefits of HPC for solving problems.
The workshop gave an overview on the advances made in these fields by the move to parallel and distributed computing. In the final plenary discussion the use and benefits of HPC were discussed from an industrial and research perspective.
The workshop, organised by GMD and NEC - the C&C Research Laboratories - drew some 70 participants, 15 from industry.
The highlight was the presence of Professor Antony James - the father of computational fluid dynamics in aircraft design and aerospace - from Princeton University. In his keynote speech on parallel CFD he discussed the issues of parallel algorithms and applications. First he noted the development of computing power and memory. From 1980 to 1985 he used an IBM 4341 with a peak performance of 0.15 Mflop/s and a memory of 2 MByte. In 1995 they installed an IBM SP2 at Princeton with 1500 Mflop/s and 6 GByte memory. That makes an improvement of 10.000 in aggregated processor speed and 3000 in memory. He estimated an average performance improvement of a factor of 20 in 5 years and of 1.8 every yearfor memory. Then he reported on cooperation and collaboration for an automatic design and pointed to the differences of fluid mechanics and structural analysis. Except StarCD the first group has inhouse and self developed codes while the second uses commercial software.
The simulation has to be done in an acceptable turnaround time at a sufficient cost and with a known level of accuracy. The accuracy of the results is in a 5% range. Dr. Jameson gave the example of the MD11. In the beginning , fuel consumption was 7% too high, 5% due to the engines. This resulted in lost orders.
The actual open issues were summarised as:
The requirements for the parallel version are performance improvement, similar ease of use as the serial code plus additional functionality. He listed the features of Star-HPC showing good parallel performance of the implicit code. Domain decomposition and message passing is integrated, as well as tools for the mesh decomposition, performance modelling and file handling. Performance goals have been achieved on benchmark problems on different platforms. The Beta-code is now available for partners.
For the automotive industry and industry in general Dr. Gosmann redefined the term MPP as Moderate Parallel Processing and not massively. He listed some application examples and the targeted platforms:
He concluded that performance improvements in real applications are rather good. In the discussion the problem was raised, will there be the same functionality for the sequential and the parallel version of StarCD at the same time? This is an open issue for all software vendors, if the parallel product is not selling so well. Will you get "older" versions ?
Shun Doi from the Central Research Lab, NEC Japan presented his experiences in interactive CFD visualisation. Nissan needs CFD computations and uses an unsteady 3D incompressible Navier-Stokes solver on a NEC SX-4/2C - Nissan's SX-4 by the way was the first shipment of that vector machine. Nissan gets a sustained performance of 1.3 GigaFlop/s.
Shun Doi summarised the design considerations for an interactive visualisation tool. The real-time visualisation simulation library RVSlib was described and parallelisation aspects discussed. RVSlib is called from a user's CFD solver running on a HPC machine. The library generates images while the calculation is running. When the job is finished, the user gets a movie directly. This tool is used by different research partners.
Professor Mark Cross from the University of Greenwich highlighted the nonlinear structural mechanics within the context of parallel, multi-physics simulation. These modes are very compute intensive, 10 to 12 variables with 10 exp 5 to 10 exp 6 meshpoints require 1000 to 10000 timesteps. Cross listed the well known challenges of parallelised code: load balancing, memory distributed equally per processor, minimisation of interprocess-communication and portability of the parallel code, Fortran 77 +PVM/MPI.
Finite volumes seem the appropriate mathematical method, as it is possible to define unstructured meshes in 3D.
Crash expert Dr. Guy Lonsdale, now with NEC, gave an overview of parallelisation paradigms for crash-simulation on HPC platforms. He summarised experiences he got from the EUROPORT project. In depth Lonsdale showed the differences between the message passing and shared memory paradigm. Surprisingly message passing has major advantages on a shared memory SGI.
Concerning the topics granularity, degree of parallelism, I/O and cache coherency he presented the differences and pros and cons of both approaches.
HPF (High-Performance Fortran) needs to be extended to deal with irregular structures. He summarised that parallel machines are a new possibility for industry and that the possibilities can be exploited - as has been demonstrated in the EUROPORT project. The scalability has to be improved by moving to dynamic mesh re-allocation. For the coupled simulation there is an increased need for HPF but there are also increased difficulties to be expected, explained Lonsdale.
Ast also listed the requirements of quality assurance - robustness in the parallel version as well as the efficiency and scalability in hardware and software and the consistency. But he noted that there is resistance of users against the parallel version. So in PERMAS the same module is used for the sequential and parallel processing. The user only has to specify the number of nodes. Ast further noted that the cache effects can be seen only with small models.
He stated: "Parallel PERMAS is standard PERMAS". His experiences show that it is scalable up to 16 nodes although the concept allows for more. Parallel computing is now ready for industrial use.
In the end two examples from forging simulation and metal forming were presented. Even these problems show good simulation behaviour. Some of the results come from the Esprit project EUROPORT D.
This third GMD-NEC Workshop at Schloss Birlinghofen took place in a creative atmosphere with a lot of discussion and information and experience exchange. The Fourth is scheduled for September next year.
© The HOISe-NM Consortium 1996