VPP5000 at AUDI AG shows good performance on CFD and crash code

 Munich 04 Apr 00 During the SAVE meeting, Karl Huber, head of CAE support at AUDI AG, described the usage of vector supercomputers and discussed the reasons for acquiring the Fujitsu VPP5000 in addition to the still existing VPP300 with 16 processors. Because of shortening the product development cycle at Audi, an enlargement of computer simulation is necessary. He also compared the performance of a RISC processor and a VPP5000 processor for typical Audi applications.

At AUDI AG the CAE system service with a small number of employees supports the engineers with computing power and software: about 60 commercial software packages; no in-house code.

In the development phase computing power and software is needed for example for linear and non-linear structural analysis, occupant safety, deep drawing, metal forming, crashworthines, kinematics, vibration and acoustic, driving dynamics and computational fluid dynamics (CFD).

At AUDI AG the team realised a decentralised computing concept with a distribution of computing resources according to the requirements of the users.

  • The CAE department has different computer systems:
  • a VPP300/16 E (16 processors 38.4 GFlop/s peak)
  • a VPP5000/4 (4 processors 38.4 GFlop/s peak)
  • a SUN E10000 (32 processors)
  • about 170 CAE sites equipped with SGI workstations, Octane - dual and single with 1 and 2 Gbyte memory
  • 7 SGI Origin 2000 with 40 R12000 processors.

The distribution of the jobs is realised by LSF (Load Sharing Facility), only the internal queues of VPP300 are managed with NQS.

Other departments have their own resources. The motivation for the acquisition of the VPP5000 was to get more computing capacity. Other reasons: there is an enlargement of computer simulations, growing demands for crash and CFD, the throughput of important analysis runs, shortening of turnaround (elapsed) time, replacement of old-fashioned computer systems and the completion of jobs with predictability in a certain time. An Audi specific benchmark demonstrated the performance differences between RISC and vector processors.

The elapsed time for StarCD V3.050 (CFD) in the HPC-version was 4130 seconds on a 16 processor SGI O2000 (R12000) compared to 4120 seconds on one VPP5000 processor - a factor of 16 on processor base. Pamcrash V98 in the 32 bit mode showed 51500 seconds on a 4 processor O2000 - more processors do not reduce the elapsed time significantly - compared to 53900 seconds on the VPP300 and about 11000 seconds on the VPP5000. This means nearly a factor of 5 compared to the VPP300 and 20 to the Origin 2000. The Nastran V70.5 results (dynamic problem, eigenvalue) are not so dramatic, approximately 4100 seconds was measured on one Origin processor, 5200 seconds on VPP300 and 1130 seconds on VPP5000.

As CFD and especially Crash demonstrate a major advantage to the user, the VPP5000 was chosen. Another argument was the good experience with the VPP300. This machine is used for crash, about 13 processors, and structural analysis, about 3 processors. The configuration of VPP5000 depends on its use: one processing element (PE) with 16 GB is dedicated to Nastran and Crash, the other 3 PEs with 8 GB memory each are used for StarCD. About 580 GB RAID storage is installed. The first experiences showed a good performance and stability. Because of the high performance of a PE it is not necessary to parallelise jobs today at Audi.

 

Uwe Harms

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