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

2022
Views:

292

Pages:

25
Author:

fluindilatxld

Broken dam Problem J.C.Martin and W.J. Moyce (1952) Initial stages of dam-break flow P.K.Stanby, A.Chegini and T.C.D.Barnes (1998) GP GPU Copyright © Takayuki Aoki / Global Scientific Information and Computing Center, Tokyo Institute of Technology 63 Experiment Simulation MULTI-GPU Performance （TSUBAME 1.2） Performance [Mflops] Number of GPUs Next Generation Weather Prediction Collaboration: Japan Meteorological Agency Meso-scale Atmosphere Model: Cloud Resolving Non-hydrostatic model Compressible equation taking consideration of sound waves. Meso-scale 2000 km Typhoon a few km Tornado, Down burst Heavy Rain Atmosphere Model Dynamical Process: Full 3-D Navior-Stokes Equation u 1  u   u    P  2  u    (   r )  g  F t  Physical Process: Cloud Physics, Moist, Solar Radiation, Condensation, Latent heat release, Chemical Process, Boundary Layer So called “Parameterization” including many empirical rules. WRF GPU Computing  WRF (Weather Research and Forecast) Community Code developed by NCAR, NCEP, OU, NOAA/FSL, AFWA WSM5 (WRF Single Moment 5-tracer) Microphysics* Represents condensation, precipitation and thermodynamic effects of latent heat release 1 % of lines of code, 25 % of elapsed time ⇒ 20 x boost in microphysics (1.2 - 1.3 x overall improvement） WRF-Chem** provides the capability to simulate chemistry and aerosols from cloud scales to regional ⇒ x 8.5 increase Initial condition Accelerator Approach Dynamics GPU Physics output Full GPU Implementation Initial condition Full GPU Approach GPU Dynamics Physics output Computatinal Flow of ASUCA CPU Initial data initialize advection GPU Long time step Coriolis force Short time step horizontal pressure gradient force Vertical pressure gradient force + gravity force（1D Helmholtz eq.） equation of continuity Update potential temperature Update pressure（EOS） output data transfer 72 Entire Porting Fortran to CUDA  Rewrite from Scratch Fortran Original code at JMA z,x,y (k,i,j)-ordering  C/C++ Changing array order x,z,y (i,k,j)-ordering CUDA GPU code x,z,y (i,k,j)-ordering 1 Year Introducing many optimizations, overlapping the computation with the communication, kernel fuse, reordering kernel execution Implementation : Advection Thread Block 64 x 4 threads (2D) in a block ny  Each thread specifies a (x, z) point、marching in y nx nz  Improve data transfer performance using domain decomposition Marching direction Using Shared Memory Shared Memory (SMem) = Software Managed Cache  Read a 2D sub-domain from VRAM into SMem   Advection : 12-point stencil  Store the xz-slice in (64 + 3) x (4 + 3) SMem 1 Block = 64 x 4 threads 64 Shared Memory 4 Access GMem directly : 4 + 4 read, 1write Using SMem read, 1write : ~1 Shared Memory VRAM (Global Memory) Access speed ~ 2 cycle 400-600 cycle Capacity 16 kByte/Block 2 GByte (Total) 2D sub-domain Halo Not in use Using Registers in marching direction GP GPU  Register Marching direction Access speed : 1 cycle  used for data not shared among threads   Shared Memory Register Advection : 12-point stencil Each thread keeps 4 y-elements in registers  Elements are reuse  Access GMem directly read, 1write 1 Thread :4+4+4 Using SMem and Registers : ~1 read, 1write VRAM (Global Mem) Implementation : 1D Helmholtz equation Thread Block ny 1D Helmholtz equation  Element in k depends on elements in k+/- 1 nx ⇒ marching in z direction nz Marching direction  64 x 4 threads (2D) in a block TSUBAME 2.0 (1 GPU) 12x speedup 50x speedup compared compared to Xeon to X5679 1 socket 1 CPU core Performance [GFlops] Performance of 5 kernels Peak Performance = 1030.0 GFlops Dynamical core (~99% of GPU time) Memory bound Fpeak = 1030.0 GFlops Bpeak = 148.0 GByte/s Physical processes (~1% of GPU time) Compute bound Arithmetic Intensity FLOP/Byte Multi-GPU : Domain decomposition z y x GPU GPU GPU GPU PCI Express (cudaMemcpy) decomposed Data exchange via MPI CPU 2D decomposition CPU CPU CPU ASUCA Typhoon Simulation 5km-horizontal resolution 479×466×48 ASUCA Typhoon Simulation 500m-horizontal resolution 4792×4696×48 Copyright © Global Scientific Information and Computing Center, Tokyo Institute of Technology 82 TSUBAME 2.0 Weak Scaling 145.0 Tflops Single precision 76.1 Tflops Doublele precision Fermi core Tesla M2050 3990 GPU SUMMARY FEATURES of GPU High Performance and Low Power ■ Major differences from Previous Accelerators ClearSpeed, Grape, , , High Memory Bandwidth suitable for wide variety of applications Consumer Product inexpensive Software Development Environment CUDA, Open CL