HiPer 2000 - High Performance Computing on Hewlett-Packard Systems

Munich 05 October 2000 More than 80 participants joined this years' HiPer (Hewlett-Packard's High Performance Computing European User Group) in Barcelona, Spain, from September 17th to 20th. It was well organised by CESCA (Centre de Supercomputacio de Catalunya - the Catalonian Center of Supercomputing). About 6 invited and 25 contributed talks had been given. The HiPer2001 will take place in Bremen from October 7 to 10, next year.

CESCA is located in the campus of the Polytechnical University of Catalonia, one of the several universities at Barcelona. Actually they operate an IBM SP2 with 44 processors, 42 thin nodes 160 and 2 wide node - a distributed memory machine, which has 12 GB main memory and nearly 500 GB disk space. The system peak performance is 27,41 Gflop/s. Hewlett-Packard is represented with two SMP (symmetric multiprocessor system) computers, an Exemplar V2500, 16 processors PA8500 (440 MHz), 8 GB memory, 216 GB disk and a peak performance of 28,16 Gflop/s. An other Hewlett-Packard computer is the N4000, 8 processors PA8500 (440 MHz), 4 GB main memory, 227 GB disk and a peak of 14,08 Gflop/s. Additionally CESCA has access to an SGI Origin2000, a Compaq Alphaserver 8400 and a Parsytec GCi, owned by The European Center for Parallelism of Barcelona (CEPBA). CESCA hosts the web site for Primeur magazine.

HiPer 2000 was divided into 3 Tutorials on Sunday, 17,

  • the CactusCode - an application framework for parallel computing
  • Pallas Tools at work - cracking performance problems
  • optimizing code for IA-64.

The opening session highlighted :"Where are you going High Performance Computing?" from different aspects. Mateo Valero, Univ. Politecnica de Catalunya, presented the processors for the future with all the new technologies to improve the processor and application performance like VLIW (Very Long Instruction Word) and EPIC (Explicit Parallel Computing).

Mathias Troyer, ETH Zuerich, gave a user's view of SMPs and PC clusters. PC clusters replace SMPs in academic settings because of better price performance and suitability for most problems - "democratisation of HPC". There will always be a coexistence, because of irregular problems, legacy codes and as nodes in high-end clusters. He additionally mentioned availability, a node has to be rebooted 0.0003/day that means, there is a limit of 3000 nodes for a day job.

Uwe Harms, HPC Consultant Munich, discussed what we can learn from the Top500 list. First there is the timely trend, an improvement of more than 50 for rank 500 within seven years. He presented three learning areas, we can learn technically:

  • efficiency, peak performanve versus Rmax
  • scaling, comparing performance gain of small and big machines of same computers
  • N1/2 - the inherent parallelism
what we can learn politically:
  • Top500 easy to understand, ranking, perhaps Rmax
  • good for simple minded politicians
  • good for high-level scientific committees
what we learned globally:
  • vendor driven and good for marketing
  • more than 100 computers commercially used, never run a Linpack
  • Linpack is not the measure to buy a system
  • trends, global results are interesting.

Fred Fuchs, Volkswagen AG Wolfsburg, gave an industrial view of HPC. A problem for industry is the simulation time. He presented an example of a side part with 436 000 nodes with 1.5 million degrees of freedom, which needs more than 14000 iterations. The old NEC SX-4 needed 14 days with one CPU, compared to the new HP machine (N-Class) with 3.5 days on 8 CPUs.

Joerg Stadler, debis Systemhaus, the IT service provider of DaimlerChrysler, presented hpcPortal, application service providing for HPC. The registered user can start his finite element job with application software like Ansys directly from his internet browser. All the data transfer between the user and the machine is encrypted, different keys are used to identify the user. He can run the job on different computers, NEC SX4/5, Cray SV1, Cray T3E, Hewlett-Packard V-class and the new Hitachi SR8000 with 16 nodes.

James C. T. Pool, Center for Advanced Computing Research at Caltech, discussed the next generation system at Caltech. Caltech has a long history of cooperation with Hewlett-Packard. Actually they use V2500 with 128 PA-8600 processors, PA-8700 are planned. Additionally they want 256 processors McKinley, but this is subject to the availability of funding. In August 14 at 1 p.m., the new Superdome, 32 processors, arrived. The installation started and on August 14 at 10 p.m. selected applications ran successfully. They measured 1.6 to 2x the performance compared to the V2500. NAS parallel benchmark class C showed a factor of 1.25 to 2.01.

Frank Baetke, Hewlett-Packard, chaired this session and the following panel discussion. The panelists presented their different views, a university of research lab has other possibilities to work with innovative computing elements and structures like clusters of workstations. Industrial users have the pressure of availability, cost and application performance. This leads to different HPC architectures for specific problems like electromagnetic compatibility or Crash simulation.

The users gave about eight talks on their experiences. Hewlett-Packard presented strategies and platforms, system and development software, transition to IA-64 and applications. The last day users and HP discussed programming and data management paradigms.

In all it was a well organised - with a lot of touristic highlights - and efficient conference with a lot of news and information.

Abstracts can be found at: http://www.hp2eug.org CESCA: http://www.cesca.es

Uwe Harms

[News on Advanced IT][Calendar][Analysis][IT in Medicine]