1 / 16

Performance study of multi-GPU acceleration of LU Factorization

Performance study of multi-GPU acceleration of LU Factorization. Paul Brandon Abbott University of Denver Dr. Yifeng Zhu University of Maine. Outline. Motivation What is the LU Factorization? Method Results Conclusion. Motivations. Multi-GPU acceleration requires setup! CUDA

kaspar
Download Presentation

Performance study of multi-GPU acceleration of LU Factorization

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Performance study of multi-GPU acceleration of LU Factorization Paul Brandon Abbott University of Denver Dr. Yifeng Zhu University of Maine

  2. Outline Motivation What is the LU Factorization? Method Results Conclusion

  3. Motivations • Multi-GPU acceleration requires setup! • CUDA • P-Threads • Communication may be required. • Synchronization. • Memory Transfers. • Overhead time is generated.

  4. Fundamental Question: Will multiple GPUs still obtain better performance than a single GPU (or CPU), given that there is overhead in communication?

  5. LU Factorization

  6. Why factor? • Solving Linear Equations becomes trivial. • The inverse of A is also easy. • Determinate of A? No problem. • Analogous to factoring a polynomial.

  7. How to factor: Step One Step Two Step One Step One Step Two …

  8. CUDA – Using a GPU • Copy memory onto device. • Specify Blocks & Threads • Execute • Copy Memory back. Steps to Use GPU Grid of Thread-Blocks

  9. Blocks have 512 threads each GPU does scheduling to handle as many blocks as possible

  10. Multi-GPU Setup: CUDA int main() { … POSIX Threads CUDA Done! CUDA CUDA

  11. The Exciting Part

  12. Conclusion • The size of working dataset influences the efficiency of multi-GPUs. • There is a time to use: • The CPU • The GPU • Multiple GPUs

  13. Thank you!

More Related