Ce diaporama a bien été signalé.
Nous utilisons votre profil LinkedIn et vos données d’activité pour vous proposer des publicités personnalisées et pertinentes. Vous pouvez changer vos préférences de publicités à tout moment.

ODCA Board Best Practice: High Performance Computing at BMW

3 512 vues

Publié le

Virtual simulations can handle more and more areas in vehicle development with better quality resulting in an strongly increasing demand for virtual simulation to complement or replace costly and time consuming physical simulations. The need for high performance computing (HPC) cycles necessary to perform these simulations grows accordingly resulting in requirements in computing power, electricity, cooling and floor space that have to be met. This talk will give an overview about our approaches to handle these demands in terms of architecture, co-location and cloud solutions.

Publié dans : Technologie

ODCA Board Best Practice: High Performance Computing at BMW

  1. 1. HIGH PERFORMANCE COMPUTING @ BMW Wolfgang Burke, BMW Group, 22.09.2014
  2. 2. AGENDA. 1 BMW Group – The company 2 HPC demands 3 HPC Cluster @ BMW 4 HPC Colocation 5 HPC Cloud Page 2 High Performance Computing @ BMW, 22.9.2014
  3. 3. High Performance Computing @ BMW, 22.9.2014 Page 3 BMW GROUP – THE COMPANY. KEY FIGURES. The BMW Group is one of the ten largest industrial enterprises in Germany. Brands: BMW, MINI, Rolls-Royce, BMW Motorrad. Sales BM W Group 2013: 76,06 bn. Euro. Deliveries of 1.963.798 cars in 2013. Deliveries of 115.215 motorcycles in 2013. 29 manufacturing and assembly plants in 14 countries. 43 sales companies worldwide. Presence in more than 140 countries. 110.351 employees at 31.12.2013.
  4. 4. Page 4 Moses Lake Spartanburg Shenyang Chennai Rayong Kulim Jakarta Araquari Rosslyn Production Plants Manaus Assembly Plants Engines / Components BMW Group Joint Venture Contract Production Regensburg Dingolfing Berlin Eisenach Leipzig Wackersdorf Landshut Munich Kaliningrad Born (2014) Steyr Graz Hams Hall Oxford Swindon Goodwood San Luis Potosí (2019) State 07/2014 Cairo BMW GROUP PRODUCTION NETWORK WORLDWIDE. 29 LOCATIONS IN 14 COUNTRIES.
  5. 5. High Performance Computing @ BMW, 22.9.2014 Page 5 BMW GROUP – THE COMPANY. POSITIVE IMPULSES BY NEW MODELS. Mini Hatch 5 doors BMW i8 BMW X6 BMW 2 Series Active Tourer BMW X4 BMW M4 Cabrio
  6. 6. AGENDA. 1 BMW Group – The company 2 HPC demands 3 HPC Cluster @ BMW 4 HPC Colocation 5 HPC Cloud Page 6 High Performance Computing @ BMW, 22.9.2014
  7. 7. High Performance Computing @ BMW, 22.9.2014 Page 7 HPC DEMANDS CHALLENGES BY INCREASING HPC USAGE 2007-07 2008-07 2009-07 2010-07 2011-07 2012-07 2013-07 2014-07 Core hours Computation time demands per month
  8. 8. High Performance Computing @ BMW, 22.9.2014 Page 8 HPC DEMANDS INCREASING NUMBER OF APPLICATION AREAS USING SIMULATION Aerodynamics Stiffness, structural dynamics Passive Safety Operating strength Powertrain Production
  9. 9. High Performance Computing @ BMW, 22.9.2014 Page 9 HPC APPLICATIONS AERODYNAMIC SIMULATION
  10. 10. High Performance Computing @ BMW, 22.9.2014 Page 10 HPC APPLICATIONS FUEL INJECTION SIMULATION.
  11. 11. High Performance Computing @ BMW, 22.9.2014 Page 11 HPC APPLICATIONS. SIMULATION OF PRODUCTION PROCESSES. Forming limit diagram of an experimental tool Simulation of the deep drawing process
  12. 12. High Performance Computing @ BMW, 22.9.2014 Page 12 HPC DEMANDS SIMULATION MODELS GETTING MORE DETAILED
  13. 13. High Performance Computing @ BMW, 22.9.2014 Page 13 APPLICATIONS ON HPC. PREDICTIVENESS IN MATERIAL BEHAVIOR.
  14. 14. High Performance Computing @ BMW, 22.9.2014 Page 14 HPC DEMANDS. SIMULATION REDUCES PROTOTYPES IN THE DESIGN PROCESS. design stage 2 design stage 1 industrialisation 1 2 3 Reduced prototype driven design process (schematically): “close to production” cars design stage 2 design stage 1 industrialisation 1 2 3 Old prototype driven design process (schematically): prototypes “design evaluation” prototypes “design validation” “close to production” cars prototypes “design validation” (only in case of high innovations) Benefits of the new process: Straightforward design process. Reduced number of prototypes and cost savings.
  15. 15. High Performance Computing @ BMW, 22.9.2014 Page 15 HPC DEMANDS CRASH SIMULATION: FRONTAL IMPACT Experimental model for method development
  16. 16. High Performance Computing @ BMW, 22.9.2014 Page 16 HPC DEMANDS. NUMBER OF LEGAL REQUIREMENTS INCREASE. Heute 1990 1997 2005 FMVSS 208 old FMVSS 208 new Phase I FMVSS 208 new Phase II FMVSS 214 static FMVSS 214 static + dynamic FMVSS 214 new MDB + Pole FMVSS 301 old rear crash FMVSS 301 new Offset rear crash FMVSS 216 roof crush test FMVSS 216 new Occupant Containm. ECE R95 side crash ECE R95 AE-MDB Advanced European Movable Barrier 2003/102/EC PDB or TRL Barrier NHTSA Front to Front BMW model USA ECE USA ECE USA ECE USA USA ECE Frontal impact Side impact Rear impact Fuel tightness Rollover Pedestrian safety Compatibility ECE R12 steering column displacement ECE R94 Offset frontal crash 2011
  17. 17. High Performance Computing @ BMW, 22.9.2014 Page 17 BMW product portfolio (status June 2014)  Increasing number of models, variants, facelifts  Diversity per model (e.g. left/right hand drive, motorization) BMW market introduction 2014 HPC DEMANDS INCREASING NUMBER OF MODELS AND VARIANTS
  18. 18. AGENDA. 1 BMW Group – The company 2 HPC demands 3 HPC Cluster @ BMW 4 HPC Colocation 5 HPC Cloud Page 18 High Performance Computing @ BMW, 22.9.2014
  19. 19. High Performance Computing @ BMW, 22.9.2014 Page 19 HPC CLUSTER @ BMW HPC SERVICE WITH AN EXTENSIBLE RESOURCE POOL. Shared, extensible resource pool Optimized computational architecture & technologies Transparent access Jobflow & workload management control Overall operations framework (e.g. installation, administration, monitoring)
  20. 20. High Performance Computing @ BMW, 22.9.2014 Page 20 HPC CLUSTER @ BMW HPC SERVICE WITH AN EXTENSIBLE RESOURCE POOL. Workload-Management HPC Cluster Engineers, Virtual Dev.
  21. 21. High Performance Computing @ BMW, 22.9.2014 Page 21 HPC CLUSTER @ BMW IMPACT OF HIGHER VIRTUAL SIMULATION USAGE 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 Floor space of HPC installation Goals of the BMW HPC decisions enable a scalable model for data center space requirements deal with the increasing electricity prices address carbon footprint caused by higher energy consumption 2008 2010 2012 2014 2016 2018 2020 2022 2024 Electricity price 2007-07 2008-07 2009-07 2010-07 2011-07 2012-07 2013-07 2014-07 Core-h per month
  22. 22. AGENDA. 1 BMW Group – The company 2 HPC demands 3 HPC Cluster @ BMW 4 HPC Colocation 5 HPC Cloud Page 22 High Performance Computing @ BMW, 22.9.2014
  23. 23. High Performance Computing @ BMW, 22.9.2014 Page 23 HPC CLUSTER @ BMW HPC CLUSTER DEVELOPMENT STEPS. HPC private cloud •On-demand resource reservation •Transparent access (self-service portal) •Standardized applications (services) •Fully automated •Zero downtime approach •Cluster capacity scalability •Maximum cluster utilization •Cluster technological adaptability •Cluster vendor independence •Cluster best price purchasing
  24. 24. High Performance Computing @ BMW, 22.9.2014 Page 24 HPC CLUSTER @ BMW HPC CLUSTER DEVELOPMENT STEPS. HPC public cloud •Flexible capacity scalability •Best price purchasing •“HPC as a service” Challenges: •Remote location •Unclear cloud model (IaaS, PaaS, SaaS, …) •Cloud integration •Data security •Organizational aspects •… HPC private cloud •On-demand resource reservation •Transparent access (self-service portal) •Standardized applications (services) •Fully automated •Zero downtime approach •Cluster capacity scalability •Maximum cluster utilization •Cluster technological adaptability •Cluster vendor independence •Cluster best price purchasing
  25. 25. High Performance Computing @ BMW, 22.9.2014 Page 25 HPC CLUSTER @ BMW HPC CLUSTER DEVELOPMENT STEPS. HPC public cloud •Flexible capacity scalability •Best price purchasing •“HPC as a service” Challenges: •Remote location •Unclear cloud model (IaaS, PaaS, SaaS, …) •Cloud integration •Data security •Organizational aspects •… HPC Co-Location •Space capacity scalability •DC technology adaptability •DC best price purchasing Challenges: •Remote location •Data security •Organizational aspects HPC private cloud •On-demand resource reservation •Transparent access (self-service portal) •Standardized applications (services) •Fully automated •Zero downtime approach •Cluster capacity scalability •Maximum cluster utilization •Cluster technological adaptability •Cluster vendor independence •Cluster best price purchasing
  26. 26. High Performance Computing @ BMW, 22.9.2014 Page 26 HPC COLOCATION ICELAND FOR HPC COLOCATION Data center operation with 100% CO2 emission free, renewable energy (geothermal and hydro) Free air cooling nearly the whole year Fixed low power cost for up to 20 years High scalability in data center floor space High bandwidth network connections available Power Usage Effectiveness (PUE) < 1,2 1 1,2 1,43 1,8 0 0,5 1 1,5 2 Average PUE BMW Data Center PUE Iceland Colocation PUE Best physical reachable PUE
  27. 27. High Performance Computing @ BMW, 22.9.2014 Page 27 HPC COLOCATION RELOCATION OF ELEMENTARY CELLS TO ICELAND. Workload-Management HPC Cluster Engineers, Virtual Dev.
  28. 28. High Performance Computing @ BMW, 22.9.2014 Page 28 HPC COLOCATION RELOCATION OF ELEMENTARY CELLS TO ICELAND. Workload-Management HPC Cluster Engineers, Virtual Dev.
  29. 29. High Performance Computing @ BMW, 22.9.2014 Page 29 HPC COLOCATION CHALLENGE LATENCY Batch jobs: Only one big bulk (result) data transfer Optimization of TCP stack parameters  Communication performance for bulk data transfer sufficient Distance Munich - Keflavik: 2850 km Latency BMW local 0,3 ms roundtrip Latency BMW – Iceland 53 ms roundtrip Transfer-Server Cluster Workload Management Storage system Job submit Result (bulk) data transfer WAN connection WAN connection
  30. 30. High Performance Computing @ BMW, 22.9.2014 Page 30 HPC COLOCATION TRANSFER TIMES AND OVERALL RUNTIMES FOR CRASH JOBS Transfer time EZ IS 0,73% EZ IS 0,86% EZ MUC EZ MUC 0,41% 0,41% EZ MUC 0,41% Problem size of crash job 100% 98% 97% 97% 110% HPC stages local stages
  31. 31. High Performance Computing @ BMW, 22.9.2014 Page 31 HPC COLOCATION INCREMENTAL EXTENSION OF HPC APPLICATIONS IN ICELAND including Crash HPC load share successful not successful open not tested November 2012 Initial application tests in Iceland HPC load share released not release not successful open not tested February 2013 Start of productive calculations in Iceland HPC load share released not successful open Current situation Only a few applications open or not released Crash
  32. 32. High Performance Computing @ BMW, 22.9.2014 Page 32 HPC CLOUD COLOCATION ODCA CARBON FOOTPRINT AND ENERGY EFFICIENCY USAGE MODEL Application of the ODCA Carbon Footprint and Energy Efficiency Usage Model Execution of an ODCA Carbon Footprint and Energy Efficiency Usage Model Proof of Concept
  33. 33. High Performance Computing @ BMW, 22.9.2014 Page 33 1148 673 2,1 0 200 400 600 800 1000 1200 1400 Coal Fired Grid Natural Gas Grid Iceland Data Center Carbon Usage Effectiveness (CUE) HPC COLOCATION DATA CENTER ICELAND: CARBON POC g/kWh
  34. 34. AGENDA. 1 BMW Group – The company 2 HPC demands 3 HPC Cluster @ BMW 4 HPC Colocation 5 HPC Cloud Page 34 High Performance Computing @ BMW, 22.9.2014
  35. 35. High Performance Computing @ BMW, 22.9.2014 Page 35 HPC CLUSTER @ BMW HPC CLUSTER DEVELOPMENT STEPS. HPC public cloud •Flexible capacity scalability •Best price purchasing •“HPC as a service” Challenges: •Remote location •Unclear cloud model (IaaS, PaaS, SaaS, …) •Cloud integration •Data security •Organizational aspects •… HPC Co-Location •Space capacity scalability •DC technology adaptability •DC best price purchasing Challenges: •Remote location •Data security •Organizational aspects HPC private cloud •On-demand resource reservation •Transparent access (self-service portal) •Standardized applications (services) •Fully automated •Zero downtime approach •Cluster capacity scalability •Maximum cluster utilization •Cluster technological adaptability •Cluster vendor independence •Cluster best price purchasing
  36. 36. High Performance Computing @ BMW, 22.9.2014 Page 36 HPC CLOUD MOVE ELEMENTARY CELLS IN A HPC CLOUD. Workload-Management HPC Cluster Engineers, Virtual Dev. HPC colocation provider HPC on premise
  37. 37. High Performance Computing @ BMW, 22.9.2014 Page 37 HPC CLOUD MOVE ELEMENTARY CELLS IN A HPC CLOUD. Workload-Management HPC Cluster Engineers, Virtual Dev. Stepwise approach Reference performance benchmarks HPC colocation provider HPC cloud provider HPC on premise
  38. 38. Cloud platform High Performance Computing @ BMW, 22.9.2014 Page 38 HPC CLOUD MOVE ELEMENTARY CELLS IN A HPC CLOUD. HW provisioning performed by cloud platform Servers File service (scratch) High speed network (Infiniband) •Gateway-Nodes •~132 Compute- Nodes: o Bignodes o Stdnodes •High Speed Cluster Interconnect •Admin Interconnect EZ-Scratch Building Block: Elementarzelle WAN connection HW administration provided by cloud platform On-demand provisioning of servers in a cell
  39. 39. Cloud platform High Performance Computing @ BMW, 22.9.2014 Page 39 HPC CLOUD MOVE ELEMENTARY CELLS IN A HPC CLOUD. •Gateway-Nodes •~132 Compute- Nodes: o Bignodes o Stdnodes •High Speed Cluster Interconnect •Admin Interconnect EZ-Scratch Building Block: Elementarzelle WAN connection 1. Step: Bare metal provisioning HW server OS provisioning HPC provisioning cloud service operation Reference for performance benchmarks
  40. 40. Cloud platform High Performance Computing @ BMW, 22.9.2014 Page 40 HPC CLOUD MOVE ELEMENTARY CELLS IN A HPC CLOUD. •Gateway-Nodes •~132 Compute- Nodes: o Bignodes o Stdnodes •High Speed Cluster Interconnect •Admin Interconnect EZ-Scratch Building Block: Elementarzelle WAN connection 1. Step: Bare metal provisioning HW server OS provisioning HPC provisioning cloud service operation 2. Step: Virtual server provisioning cloud service operation HPC provisioning Virtual server provisioning Reference for performance benchmarks
  41. 41. High Performance Computing @ BMW, 22.9.2014 Page 41 HPC CLOUD MOVE ELEMENTARY CELLS IN A HPC CLOUD. Seamless integration of HPC cloud into BMW HPC cluster First test results on bare metal show comparable performance to current cluster Basis for a HPC service specification Open questions to be clarified
  42. 42. High Performance Computing @ BMW, 22.9.2014 Page 42 HPC CLOUD DEVELOPMENT OF A HPC CLOUD MODEL. ODCA usage models might be helpful on different layers •Data Security Security Workgroup •Long Distance Migration Manageability & Services Workgroup •ODCA Interop Across Clouds Guide •Compute Infrastructure as a Service Infrastructure Workingroup
  43. 43. Page 43 THANK YOU VERY MUCH FOR YOUR ATTENTION Imprint: Published by: BMW Group Corporate and Governmental Affairs 80788 Munich No reproduction, in whole or in part, without the express permission of Bayerische Motoren Werke Aktiengesellschaft, München ©2014 Bayerische Motoren Werke Aktiengesellschaft

×