The Use of Clusters for Engineering Simulation: A Personal History of Clusters. Lynn Lewis Hewlett Packard CAE Segment Manager

Size: px
Start display at page:

Download "The Use of Clusters for Engineering Simulation: A Personal History of Clusters. Lynn Lewis Hewlett Packard CAE Segment Manager"

Transcription

1 The Use of Clusters for Engineering Simulation: A Personal History of Clusters Lynn Lewis Hewlett Packard CAE Segment Manager

2 A History with lessons Solving problems using the tools at hand Limits of existing computing resources Proprietary systems Getting around limitations Commercial focus & ROI Patterns that work Unix (Linux - look and feel, smell and taste circa 1983) Will open source work for enterprise and the critical path Can hardware companies make money why you should care Value proposition Cost of labor (Engineers cost money) Value to user Peter Schumacher, Project Overview BMW Group

3 CDC ETA 10 FSU: Four independent systems Operating System No tools Broke Single processor worked fine Hydrodynamic stability Series solutions Perturbation methods Peter Schumacher, Project Overview BMW Group

4 Digital Computer and Cray VMS Cluster Eglin AFB Armament Laboratory symmetric and two dimensional flow curvature terms in boundary layers CRAY 1 & 2 Krikland AFB Electronics Laboratory Short on memory Solid state device (SSD) fast disc-like blocks of memory Computational Fluid Dynamics EAGLE (manual domain decomposition) Finite Element HULL (serial then ported to SMP) Peter Schumacher, Project Overview BMW Group

5 SGI - MIPS Fermi Lab A couple hundred workstations (particle physics) Special no keyboards, mice, and NO Graphics (horror) SMP (symmetric multi-processor) Accidental Irix (the legend of Rocky Roads ) Software and hardware context switching Graphics Two processors, four processors, eight processors Peter Schumacher, Project Overview BMW Group

6 SGI-MIPS Defense Intelligence Agency Huntsville, AL Signal processing (SMP) Background: The Grench that stole Christmas three times The Grench that saved SGI the first time Crimson I want Red! IBM (clever clusters SP1) + Grumman (disgruntled integrator) NRC & SGI (Cray + SGI clustered systems) Peter Schumacher, Project Overview BMW Group

7 PVM MPI Open MP (standard of the day) PVM Peak usage around 1995 MPI (1, 2) Still growing, not falling off Clusters and SMP OpenMP Growing with a lot of interest SMP and Clusters sort of Peter Schumacher, Project Overview BMW Group

8 ESI-Group + (a brief case study) ESI PAM, Crash (Germany and Japan) + CD-Adapco, CFD (Germany and Japan) Investment from Euro-port Investment from IBM (SP1 SP2) Investment from Cray (T90) merger with SGI (investment funds for support) Domain Decomposition Crash Commercial Product Likewise Mecalog (Radioss), LSTC (Dyna) Fluent, and MARC Working at Ansys, Abaqus, MSC.Software Peter Schumacher, Project Overview BMW Group

9 Major European Automotive Build & Design Crash, Crash and more Crash: or how hardware companies learned to love non-deterministic methods Large Visualization system small cluster x cluster How it worked How it works Peter Schumacher, Project Overview BMW Group

10 Motivation - ROI ROI Faster time to market Fewer Engineering Change Orders More reliable products Fewer Cadavers More time at a value proposition (higher prices) More facelifts, less time Government Regulations Liability Peter Schumacher, Project Overview BMW Group

11 Automotive industrial goals Potential cost saving through faster design Vehicle development time to be reduced from 38 to 30 months in the USA from 30 to 26 months in Japan Neon Escort Months development time Vehicle virtual engineering to increase from 40% up to 70% Vehicle physical prototypes to decrease by 40% Saturn Neon time in months Sources: Wards Auto World, 11/93. Pp 381 Design Management Journal, Summer 1998, p 50 development expense Source : OSAT University of Michigan Forecast and Analysis of the North American Automotive Industry [for 2009] Escort Saturn $B development expense

12 Technological Innovation Computing Cost -29% Annually Courtesy GM, and Steve Rohde, now at Quantum Signal, LLC

13 Accelerated Technological Innovation: Radical Impact 7 10,000,000

14 Radical Impact: Substitution Effects Or, the inverse, Engineers are 10,000,000 X s more expensive per unit of computing at 2001 costs. Relative Production Cost Factors shifted by 1 10,000,000 Computers are now free compared to the Engineer at 2001 Costs

15 Answer to industrial needs price/performance Front impact model Cost of one run Machine amortization + operating cost over 3 years) 18,0 Courtesy: BMW AG 16,0 14,0 12,0 10,0 2000D Speed-up 2002D Speed-up Ideal speed-up 8,0 6,0 4,0 2,0 0, x better cost 15x better speed Processor speed: only 1.8x element frontal crash Scalability of DMP versions

16 Answer to industrial needs Turnaround time Time reduced despite increasing model complexity Model construction still 2 weeks Simulation time overnight 2,000,000 1,500,000 Number of elements in crash model 2,000,000 1,000, , ,000 0 VW 5, ,

17 Other Examples IA32 with Domain Decomposition and Segmented Memory Fluent Fluid Dynamics LSTC Dyna Crash & Stamp

18 Commercial Linux IA32/IA64 (specific players and configurations) MSC.Software LSTC ESI-PAM Mecalog Fluent CD-Adapco Abaqus Ansys

19 Conclusions on Performance Engineering TIME is critical to success! Engineering methods: automatic grid generation resulting in larger grids, and more time steps for stability require 64 bit floating point calculations and addressing (>= 200K elements) Current IA64 is >2x the speed of first generation IA64 systems with only a 25% increase in clock frequency Current parallel performance is 2x faster than parallel performance on first generation systems clustering capability of analysis tools enable large cost effective systems to be constructed Hierarchical memory architecture effectively feeds the beast for more information:

20