Fast & Furious: building HPC solutions in a nutshell
1 like1,260 views
Victor Haydin presented on high-performance computing (HPC). He began by defining HPC as using supercomputers and computer clusters to solve advanced computation problems, such as those involving teraflops-level performance. He discussed why HPC is needed for applications in areas like finances, healthcare, fluid dynamics, and genetics. Finally, he outlined some approaches to implementing HPC, such as using commodity hardware, GPU-based systems, distributed and load-balanced architectures, and middleware to handle errors and allow the same code to run on CPUs and GPUs.
Fast & Furious: building HPC solutions in a nutshell
- 1. Victor Haydin Head of R&D, ELEKS
- 2. Agenda 1. What is HPC? 2. Why does somebody need it? 3. How to do it?
- 3. What?
- 4. Definition Wikipedia: “High-performance computing (HPC) uses supercomputers and computer clusters to solve advanced computation problems. Today, computer systems approaching the teraflops- region are counted as HPC-computers.”
- 5. Definition advanced computation problems
- 8. Big Data
- 9. A.I.
- 10. Supercomputers Computer clusters Teraflops performance
- 11. HPC systems comparison 100000000 10000000 1000000 100000 10000 HPC 1000 100 10 1 CPU (Intel Ivy Bridge) 100xCPU GPU (NVIDIA Kepler) 100xGPU IBM Sequoia
- 12. Why?
- 13. Finances
- 14. Healthcare
- 16. Genetics
- 17. Computer Vision and Image Processing
- 18. How?
- 19. Disclaimer
- 21. Specialized VS.
- 22. GPU-based
- 23. Example 1: Financial Risk Analysis Using Monte-Carlo method On GPGPU
- 28. Distribute
- 29. Run
- 30. Define
- 31. Store
- 32. Feed
- 33. Present
- 34. Survive
- 36. Middleware
- 37. Worker
- 38. Example 2: Image Search platform Using local feature detection On GPGPU
- 53. Middleware
- 54. Load Balancing
- 57. Unicast 140 • Computation time – 1 second 120 • Sending time – 120 seconds! 100 80 Unicast • More workers – slower speed 60 40 20 0 9 workers 18 workers
- 58. Multicast 140 • Computation time – 1 second 120 • Sending time –25 seconds 100 80 Unicast Multicast • Almost 5 times faster 60 40 20 0 1 2
- 59. Middleware
- 60. Worker
- 62. ERROR: CUDA ERROR CODE 30 (“UNKNOWN ERROR”)
- 63. Run same code on CPU and GPU
- 64. Kernel CUDA_KERNEL foo(…) { CUDA_DEFINE_PARAMS; // your code here } CUDA_CALL(threads, blocks, foo(…))
- 65. Generated code // GPU mode // CPU mode __global__ void foo (…) void foo(…) { { // same code here // your code here } } // LOOP OVER threads and blocks foo<<<threads, blocks>>>(…) { foo(…) }
- 66. Pros & Cons • Same code for CPU and • Shared memory GPU • __syncthreads() • Debugging • Range checking • No CUDA ERROR 30
- 69. Got a question? Ask!