University:
Boston University
Course:
CAS CS 320 | Concepts of Programming Languages
Academic year:

2020
Views:

378

Pages:

20
Author:

mondaroslz1u

NVIDIA GPU GPU Memory Bpeak/Fpeak Peak Performance [GFlops] Intel Core i7 Extreme Tesla C2050 /M2050 51.2*,102.4 515*,1030 Number of Processor 4 GeForce GTX 580 Fermi 197*,1576 448 512 Core Clock [MHz] 3200 1476 1544 Bandwidth[GB/s] 32 148.8 192.1 Memory Interface [bit] 64 384 384 Memory Clock [GHz] 1.333 (DDR3) 1.55 (GDDR5) 2.00 (GDDR5) ----0.624 3.0 0.289 1.536 0.974 Capacity [GB] Bandwidth/Performance Tesla M2050 Peak Power : 225W Peak Power : 244W GPU Architecture On Chip GF100, GF103, GF104 Register Memory Video Memory Global memory Streaming Multiprocessor Shared memory + L1 Cache Streaming Processor (CUDA core) Off Chip ~6GB (VRAM ~16 (C2050 (GF100): 14) 64 Kbyte 8～48 per SM, total 512 Heterogeneous Computer ■ Several Bandwidth Bottle Necks Memory DDR3-1600 32 GB/sec CPU GeForce GTX 580 GPU Memory Interface 384bit 192 GB/sec 1000BASE-T 0.125 GB/sec InfiniBand QDR 40 Gbps (5 GB/sec) VRAM PCI Express 2.0 x16 80 Gbps (10 GB/sec) CFD Performances in GPU Computing Partial GPU Implementation 30% up ～ ×3 ※ Only Hot spot (Intensive part) : small cost ※ Overhead of host (CPU) memory device (GPU) memory communication FULL GPU Implementation ×10 ～ ×100 ※ Limitation of device (GPU) on board memory size Real-time TSUNAMI Simulation ADPC : Asian Disaster Preparedness Center Early Warning System: Data Base Real‐time CFD high accuracy Shallow‐Water Eq. Conservative Form: Assuming hydrostatic balance in the vertical direction, 3D 2D equation h hu hv   0 t x y z hu   2 1 2  huv   hu  gh     gh 2 x t x   y z hv huv   2 1 2     hv  gh    gh t x y  2 y  Tsunami Modeling free surface flow Water Height : h H Bottom Height : z Directional-Splitting Method U F G   S t x y hv  hu  h     2 1 2   U  hu  , F  hu  2 gh  , G  huv  hv 2  1 gh 2  huv  hv    2   First Step: x-directional computation h hu  0 t x hv uhv  0 t x hu   2 1 2  z   hu  gh    gh t x  2 x  Second Step: y-directional computation h hv  0 t y hu vhu  0 t y z hv   2 1 2    hv  gh    gh 2 y t y   For Characteristics-based Method For Conservative Semi-Lagrangian Method CIP-CSL2 (Conservative Semi-Lagrangian) R. Tanaka, T. Nakamura, and T. Yabe, Comp. Phys. Comm., 126, 232‐243 (2000). hi hi 1 Continuum Eq. h i h hu  0 t x  t  t t h iX1/ 2   hi ( x)  a ( x  xi )  b( x  xi )  hi 2 hx ,i  6h x  xi h ( x) dx udt x p xi 1 xi X i 1 / 2 2 xi  x h i 1 2hi 1  4hi x 3hi 1  3hi 6hiX1/ 2 6hiX1/ 2 2hi 1  4hi  , b a  2 3 2 x x x x hi ( xi )  hin hi ( xi  x)  hin1 hin 1  h nj ( xi  ut ) hiX1, /n21  hiX1,n/ 2  hin1  hin  x i  x xi hiX1   xi 1 xp hi ( x )dx  hiX1/ 2 hin ( x)dx  a n 3 bn 2        hin   2  3  2-dimensional Variable Configuration X i 1 / 2 , j 1 h hi , j 1 Y i , j 1 / 2 h hi , j XY i 1 / 2 , j 1 / 2 h X i 1 / 2 , j h hi 1, j 1 hiY1, j 1/ 2 hi 1, j Characteristics-Based Method hu  U F h    0 U   , F   2 1 2  hu  gh  t t x hu  2   h n 1 u n 1 t n 1 Factorizing: C 1  F 0 U U  2 A 0 A  U   u 2 2u  t x for   gh C   tn i‐1  u C t i C  t  u i+1 Riemann invariants : 1     u  0  W W 2 ,   u    W  0   0    u 1 x t     u   2  1 1   n 1 1      u  n 1  n 1       u  u  2  2 2     1    n 1  0 u  u  u   2    t x 2      h n 1 (hu ) n 1 Hydrostatic Balance (1/2) H=h+z h=Hz z hu   2 1 2    hu  gh    gh 2 x t x   h z hu   2 1 z    hu  g ( H  z ) 2    g ( H  z ) t x  2 x  1 z z hu   2 1    hu  gH 2  ghHz  gz 2    gH  gz 2 2 x x t x   z H z hu   2 1    hu  gH 2   gH  gz  gH 2 x x x t x   H hu   2 1    hu  gH 2   gz t x  2 x  Hydrostatic Balance (2/2) For characteristics-based method, h hu  0 t x H hu   2 1    hu  gH 2   gz 2 x t x   hH z H Hu zu   x x t zu Hu   2 1    Hu  gH 2   u x  2 x t  Characteristics-based Method for H and Hu + Fractional Step : H zu  x t Hu zu u t x CUDA GPU Computing x‐directional Computation 2*blockDim.x 1 Data read n‐step ny Calculation Data write to n+1 step nx CUDA GPU Computing y‐directional Computation blockDim.x=16 2*blockDim.y Data read & write ny Calculation nx SCREEN Capture ×62 Core2 Quad 1 Core Large-scale Real-time Tsunami Simulator GPUs 384x384x192 384x384x192 384x384x192 8 GPU 400km×800km （100m mesh） 384x384x192 GPU No. within 3 min Multi-GPU : Domain decomposition 2D decomposition GPU (1) GPU→CPU CPU Node 1 Node 2 PCI Express (cudaMemcpy) (2) CPU → CPU MPI x (3) CPU→GPU Overlapping between Computation and Communication Asynchronous Data transfer SendBuffer Sync Asynchronous Data transfer MPI RecBuffer Stream 1 Stream 2 Large-scale Real-time Tsunami Simulator 100 10 8 GPU 400km×800km （100m mesh） 1D Decomposition 2D Decomposition Ideal 1 1 10 GPU No. 100 within 3 min CPU-GPU Performance Comparison (1 CPU Core based) GPU‐CPU Scalability 10000.00 10 GPU = 1000 CPU Core Speed Up (x) Speed Up 1000.00 CPU 100.00 GPUs ×66 10.00 1.00 1 10 Devices Number ofNumber GPU/CPU core 100 1000

Tsunami Simulation on NVIDIA GPU

Related Documents

Recommended Documents

Description

Related Documents

Free up your schedule!

Take 5 seconds to unlock