1. Societal applications of large scalable parallel computing systems ARTEMIS & ITEA Co-summit, Madrid, October 30th 2009

2. 2©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Outline What are large scalable parallel systems today? What is at stake with these systems? What are some of the technology challenges for tomorrow?

3. 3©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Example of an HPC system 10,000s of processors High speed interconnect Very fast IO system Petabytes of storage Applications ranging from 100s to 100,000s of cores

4. 4©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Example of an Internet data centre 10,000s of servers Standard networks Moderate to large databases Large data centre infrastructures Serving 100,000s users

5. 5©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Container data centre

6. 6©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Comparison In common - Large number of processors - Large infrastructure - Density - Energy and cooling Differences - 64bit floating point - Fast interconnect - Resilience - Few users - More diverse processing - Few communications - Restarting a new user session - Large number of users

7. Why?

8. 8©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Oil and gas : discovery Seismic acquisition campaigns Large set of data Simulation is a must for - Decision to drill - Reducing data acquisition spending Discovery of a new 50 Mboe represents ~5 B$

9. 9©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Oil and gas : reservoir simulation Today oil fields are not fully exploited in average only 25% - 30% of the oil is extracted Simulation can help for - Location of drill - Pumping dynamics - Injection of fluids 1% of the oil contented in today fields represents the annual consumption of oil worldwide

10. 10©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Automotive industry Simulation has been key for security improvement Today state of the art crash simulation has replaced test ; only one test for certification Objectives : - Reduce weight while maintaining security - Optimization of energy - Electromagnetic simulation - Acoustic simulation

11. 11©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009

12. 12©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009

13. 13©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009

14. 14©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009

15. 15©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Medicine Today simulation plays only a minor role But large effort on going - Simulation of human body - Different scales - Distributed effort worldwide - Long term effort - Surgery - Preparation of intervention - Prosthesis design - Customized treatment

16. 16©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Example of customized treatment

17. 17©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Digital economy Web economy - Web ad revenue ~$20B Internet commerce - B2B in Europe : 2,000 B € - B2C in Europe : 500 B € Productivity improvement - 25% of worldwide growth

18. 18©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 The cloud computing paradigm Transforming IT technologies into an universal accessible service Offering - Infrastructure - Development platform - Software - Business service as a service which is - Elastic - Scalable - Paid according to use Shaping a new way for digital economy

19. Challenges?

20. 20©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Processor evolution : power limit # transistors Frequency Watt Core

21. 21©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Optimization is key for progress Power cost is already a dominant factor. Its importance will keep growing. Every stage of power usage has to be optimised - Electronics design : Strained silicon or SOI, Transistor leakage control techniques, clock gating, policy based power allocation - System architecture - System Power supply and delivery : fine grain power management - Cooling system : vertical integration of cooling solutions

22. 22©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 GPU technology

23. 23©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 What GPU brings More computing power per watt - ~800 Gflops per 200W versus 70 Gflops per 100W Cheap computing power - 1 gflops/$ versus 0.14 Gflops/$ Integration challenges Heterogeneous computing More parallelism : more than a factor of 100

24. 24©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Parallel programming How to design a parallel scalable code up to million threads? What are today tools?

25. 25©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Timeline Initialization Computation: 1:00 min

26. 26©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Zooming computation: Timeline

27. 27©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Zooming computation: Process Timeline 724 MFlop/s average

28. 28©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 One computation step: Process Timeline

29. 29©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Interesting direction : hierarchical programming What does it means - Different scales in application - Different sets of data Why it is may be a solution - Fits well with human brain : abstraction, general/detailed view - Fits well with architecture : processor, node, interconnect trend How to go in this way - Already possible with a mix of OpenMP / MPI - Integration of OpenCL is needed

30. 30©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Conclusion for Artemis and ITEA Large parallel computing centres are a key technology for - Innovation - Digital economy development - Societal challenges (aging, global warning … ) Design and development of scalable software is a must for getting the most out of these centres Design and development of parallel software is today the bottle neck for large adoption of simulation

31. 31©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009

32. 32©Bull, 2009Artemis & ITEA Co-Summit, Madrid 2009 Credits Many thanks to - CEA/DAM - France3 - Google - Renault - Total - CEA/DSV - Nvidia - Dresden University