Parallel Animated Image File Generation Nishad Patel Department of - PowerPoint PPT Presentation

Parallel Animated Image File Generation Nishad Patel Department of Computer Science Rochester Institute of Technology May 2012

Introduction • Animated images • Parallel generation

Agenda • Background • Framework • Hypotheses • Simulations • The File Format • Performance • Conclusions

Background Parallel Simulation and Visualization Modern-day simulations • Limited resources • Large data – Study offline • Simulate over time – moving images • Difficulty - performance

Background File formats • Lossless compression • Few formats exist – File size vs. Data Integrity • None support parallel generation • MNG – Multiple-image Network Graphics – Lossless – High compression (Network-oriented) – No support for parallel generation

Framework • PJGMovie – int[][] matrix – PJGColorFrame • 24-bit Run-length encoding – PJGHueFrame • 24-bit Huffman encoding – PJGGrayFrame • 8-bit Huffman encoding

Framework • write(frame) • writeColSlice(Range, frame) • writeRowSlice(Range, frame) • writePatch (Range, Range, frame) • getPixelDataSegmentParameters() • readSegment()

Hypotheses • Serial PJM processing times vs. Parallel PJM processing times – Parallel PJM should be quicker • Post-processing times vs. In-situ processing times – In-situ processing should be quicker • PJM processing times vs. processing times for other formats – PJM processing times will be smaller • PJM file sizes vs. file sizes for other formats – PJM file sizes expected to be larger

Simulations Belushov-Zhabotinsky Reaction • Temporally oscillating chemical reaction • Reaction-diffusion system • Three substrate model • Solution in 2-D plane • Equations defined in (Turner, 2009)* A + B -> 2A B + C -> 2B C + A -> 2C *Turner, A. (March 2009). A Simple Model of the Belousov-Zhabotinsky Reaction from First Principles. Implemenation Note, University College London, Bartlett School of Graduate Studies, London

Simulations Belushov-Zhabotinsky Reaction • Reduced to a t+1 = a t + a t (b t − c t ) b t+1 = b t + b t (c t − a t ) c t+1 = c t + c t (a t − b t ) • Diffusion effect From Wikipedia en.wikipedia.org/wiki/File:The_Belousov-Zhabotinsky_Reaction.gif

Simulations Belushov-Zhabotinsky Reaction Parallelizing the program • Split frames into slices

Simulations Antiproton n-body simulation • All antiprotons have equal negative charges • Repulsive force between two antiprotons – Directly proportional to product of charges – Inversely proportional to distance between them • Perpendicular magnetic field – Forcing circular motion • Also in 2-D space

Simulations Antiproton n-body simulation Position calculation: v' = v + δ a p' = p + δ v + 1/2 ( δ * δ ) a

Simulations Antiproton n-body simulation Parallelizing the program • Split the antiprotons over the processes • Each process stores – Entire position array – Slice of velocity array – Slice of acceleration array • All-gather

File Format • Header segment 0x00, 0x50, 0x4A, 0x4D, 0x00, 0x00, 0x00, 0x01 • Image type segment • Height segment • Width segment • Number of Frames segment • Frame rate segment • Creation time segment • Comment segment • Run length encoded 24-bit color pixel data segment • Huffman delta encoded 8-bit grayscale pixel data segment • Huffman delta encoded 24-bit hue pixel data segment

Performance Serial PJM processing times vs. Parallel PJM processing times B-Z Reaction 200 x 200 400 x 400 NT T Speedup NT T Speedup Serial 17753 Serial 78486 1 19280 0.921 1 68876 1.140 2 9225 1.924 2 35217 2.229 3 6433 2.760 3 24214 3.241 4 5160 3.441 4 18713 4.194 8 3596 4.937 8 10541 7.446 Serial Post-processing vs Parallel Post- Serial Post-processing vs Parallel Post- processing processing 100000 Running Times (msecs) Running Times (msecs) 10000 400 x 400 200 x 200 10000 1000 1 10 1 10 Number of Threads Number of Threads

Performance Post-processing times vs. In-situ processing times 200 x 200 B-Z Reaction 400 x 400 NT T TT Speedup NT T TT Speedup 1 16129 43734 2.712 1 74079 175886 2.374 2 8634 22036 2.552 2 36071 88119 2.442 3 6206 15335 2.471 3 24436 60197 2.463 4 4982 12198 2.448 4 18848 46599 2.472 8 3299 7906 2.396 8 11141 34181 3.068 Post-processing vs Insitu-processing Post-processing vs Insitu-processing 100000 Running Times (msecs) Running Times (msecs) 100000 Post-processing 10000 Insitu-processing 1000 10000 1 10 1 10 Number of Threads Number of Threads

Performance Post-processing times vs. In-situ processing times Antiproton Simulation NT T TT Speedup Serial 200182 202010 1.009 400 Antiprotons 1 194127 192680 0.993 2 118412 120321 1.016 3 98029 100449 1.025 4 84279 85547 1.015 Post-processing vs Insitu-processing 8 77303 78574 1.016 Running Times (msecs) Post-processing Insitu-processing 50000 1 10 Number of Threads

Performance PJM processing times vs. processing times for other formats B-Z Reaction 200 x 200 400 x 400 NT T Speedup NT T Speedup MNG File 22527 MNG File 108769 1 16129 1.397 1 74079 1.468 2 8634 2.609 2 36071 3.015 3 6206 3.630 3 24436 4.451 4 4982 4.521 4 18848 5.771 8 3299 6.828 8 11141 9.762 PJM vs MNG PJM vs MNG 100000 Running Times (msecs) 10000 PJM MNG 1000 10000 1 10 1 10 Number of Threads Number of Threads

Performance PJM processing times vs. processing times for other formats Antiproton Simulation NP T TM Speedup 200 Antiprotons 1 54058 175979 3.255 2 41873 161936 3.867 3 39719 161859 4.076 4 41351 163864 3.963 8 53834 177052 3.289 PJM vs MNG Running Times (msecs) 100000 MNG PJM 10000 1 10 Number of Threads

Performance PJM file sizes vs. File sizes for other formats MNG bytes PJM bytes Larger by % 200 x 200 pixels 20901545 22578512 8.023 400 x 400 pixels 83593947 91148715 9.037 PJM vs MNG 100000000 90000000 80000000 70000000 File size (bytes) 60000000 50000000 MNG 40000000 PJM 30000000 20000000 10000000 0 200 x 200 400 x 400 Size of image

Performance PJM file sizes vs. File sizes for other formats MNG bytes MNG bytes PJM bytes Larger by % Larger by % Antiprotons Compression 1 Compression 9 Compression 1 Compression 9 200 420515 116927 241407 -42.593 106.457 400 451694 129942 277286 -38.612 113.921 PJM vs MNG 500000 450000 400000 350000 File size (bytes) 300000 250000 MNG 1 200000 PJM 150000 MNG 9 100000 50000 0 200 400 Number of Antiprotons (N)

Conclusions • Serial PJM processing times vs. Parallel PJM processing times – Parallel PJM is quicker • Post-processing times vs. In-situ processing times – In-situ processing is quicker • PJM processing times vs. processing times for other formats – PJM processing times are smaller • PJM file sizes vs. file sizes for other formats – PJM file sizes expected are larger - generally

Future Work • Delta encoding • Player features

Thank you. • Questions?