Speedups of Inspiral Searches with GPUs

1
Speed-ups of Inspiral Searches with GPUs
Shin Kee Chung Linqing Wen
2
In Collaboration with

• Amitava Datta and David Blair (UWA)?
• The Caltech LIGO group, special thanks to Chad,
  Kipp, Drew, Alan, Phil, and Stuart
• Shin Kee Chung is supported in part by the
  Caltech SURF program

3
Hardware Used

• Timing for inspiral searches in the CPUs are
  done on a Dell Inspiron 530 computer with a 2.5
  GHz Intel Core 2 Quad 9300 CPU, and 4 GB of RAM.
• Several Graphics Processing Units (GPU) cards are
  tested. Results from Nvidia GeForce 8800 Ultra
  are presented

4
Method

• GPU programs are written in a C-like language
  CUDA. Use CUDA FFT library
• Use lalapps_inspiral with fixed parameters
• Performance comparisons
• speed-ups of CUDA-FFT vs CPU-FFT (FFTW)
• speed-ups of lalapps_inspiral with FFTW replaced
  with CUDA-FFT
• speed-ups of lalapps_inspiral chi-square test
• Use CUDA FFT, bundle up CUDA-FFT for one loop
• Implement data parallelism for the 2-loops where
  chi-square are calculated

5
Comparison of CUDA-FFT and FFTW
1 million data points 5x speed-up 4 million
data points 8x consistent with known benchmarks
6
Up to 4x Speedup is Achieved For Inspiral Search
using CUDA FFT
- lalapps_inspiral uses a fixed 1-million point
FFT - more speed-ups are expected for more data
points
7
16x Speedup is Achieved with Chi-square
Implementations
6-hrs CPU time can be reduced to 20 mins with a
GPU
8
Conclusion and Future Works

• 16x speedup can already be achieved with a small
  modules being implemented.
• The Chi-square calculations still occupies about
  80 of total execution time of the
  lalapps_inspiral for analysis of about 700
  templates. Therefore we aim to further optimize
  the Chi-square calculations.
• Suggestions are welcome