Free In Situ Volume Compression Using NVENC Nick Leaf, Bob Miller, - PowerPoint PPT Presentation

“Free” In Situ Volume Compression Using NVENC Nick Leaf, Bob Miller, and Kwan-Liu Ma UC Davis

A supercomputer is a device for turning compute-bound problems into I/O-bound problems. -Ken Batcher

Video Processing Unit (VPU) Dedicated ASIC • • Energy efficient Widely available • NVENC • Kepler (Titan) and later • Special-purpose API • http://on-demand.gputechconf.com/gtc/2017/presentation/s7111-abhijit-patait-nvidia-video-technologies.pdf

Y Y Z Time X X

Proving Usefulness Quality Compression ratio Performance

Methodology Seven volume datasets • Multiple compressors • a. Libx264, Libx265, NVENC h264, NVENC h265 Quantitative Volume Statistics Pixel Format Encode Decode Render Qualitative

Ground Truth Examples Supernova NCAR Plume Argon Bubble

Conversion Results: Quantitative Conversion Only Conversion + Compression int8 int16 int24 int32

Conversion Results: Qualitative Raw int32 Raw int8 Compressed int8

Lossiness Comparison: Quantitative

Lossiness Comparison: Qualitative QP 15 QP 30 QP 40

Proving Usefulness Quality Compression ratio Performance

Compression Ratios

Proving Usefulness Quality Compression ratio Performance

Integration with HPGMG-CUDA High-Performance Geometric Multi-Grid solver benchmark • Multi-level hybrid CPU/GPU Finite-Volume (FV) solver • https://hpgmg.org • CUDA version: https://bitbucket.org/nsakharnykh/hpgmg-cuda • Why did we choose HPGMG-CUDA? • • Worst-case encoding target = strongest case for technique Compression integration • One new dependency: libnvidia-encode.so • • Less than 100 lines in application, plus helper code

In Situ Results

Raw Data Load vs Load + Decode

Proving Usefulness Quality Sufficient for Vis Compression ratio 100:1 or better Performance Tiny in situ impact (“Free”)

Acknowledgements Sponsored in part by the U.S. Department of Energy via grants DE- • SC0007443 and DE-SC0012610 under program manager Lucy Nowell. Thanks to NVIDIA for accepting my talk! • For more details, see Leaf, Nick, Bob Miller, and Kwan-Liu Ma. "In situ video encoding of floating-point volume data using special-purpose hardware for a posteriori rendering and analysis." In 2017 IEEE 7th Symposium on Large Data Analysis and Visualization (LDAV) , pp. 64-73. IEEE, 2017.

Thank you!

Datasets Name Dimensions Min Max Argon Bubble 256x256x640 1 2.67 JHTDB QCR 1024 3 -1.76E4 5.99E4 Marschner-Lobb 512 3 1.18E-1 8.82E-1 NCAR Plume 252x252x1024 2.08E-6 6.5E1 Random 512 3 1.16E-8 1 864 3 Supernova 2.02E-15 1.25E-1 Visible Female 512x512x1734 0 4.03E3