porting the physical parametrizations on gpu using directives n.
Download
Skip this Video
Loading SlideShow in 5 Seconds..
Porting the physical parametrizations on GPU using directives PowerPoint Presentation
Download Presentation
Porting the physical parametrizations on GPU using directives

Loading in 2 Seconds...

play fullscreen
1 / 23

Porting the physical parametrizations on GPU using directives - PowerPoint PPT Presentation


  • 61 Views
  • Uploaded on

Eidgenössisches Departement des Innern EDI Bundesamt für Meteorologie und Klimatologie MeteoSchweiz. Porting the physical parametrizations on GPU using directives. X. Lapillonne, O. Fuhrer. Outline. Physics with 2d data structure Porting the physical parametrization to GPU using directives

loader
I am the owner, or an agent authorized to act on behalf of the owner, of the copyrighted work described.
capcha
Download Presentation

PowerPoint Slideshow about 'Porting the physical parametrizations on GPU using directives' - waldo


An Image/Link below is provided (as is) to download presentation

Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author.While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server.


- - - - - - - - - - - - - - - - - - - - - - - - - - E N D - - - - - - - - - - - - - - - - - - - - - - - - - -
Presentation Transcript
porting the physical parametrizations on gpu using directives

Eidgenössisches Departement des Innern EDIBundesamt für Meteorologie und Klimatologie MeteoSchweiz

Porting the physical parametrizations on GPU using directives

X. Lapillonne, O. Fuhrer

outline
Outline
  • Physics with 2d data structure
  • Porting the physical parametrization to GPU using directives
  • Running COSMO on an hybrid GPU-CPU system
new data structure
New data structure
  • 2D data fields inside the physics packages with one horizontal and one vertical dimensions: f(nproma,ke), with nproma = ie x je / nblock.
  • Goals:
    • Physics package could be shared with ICON code
    • Blocking strategy: all physics parametrization could be computed while data remains in the cache
  • organize_physics should be structured as follow:

call init_radiationcall init_turbulence …

do ib=1,nblock

call copy_to block

call organize_radiation

call organize_turbulence

call copy_back

end do

  • Note : an omp parallelization could be introduced around the block loop

where data inside organise_scheme is in block form t_b(nproma,ke)Routines below organize_scheme will be shared with ICON. Fields are passed via argument list:

call fesft(t_b(:,:), …

current status
Current status
  • Base code: COSMO 4.18
  • 2d version of microphysics (hydci_pp), radiation (Ritter-Geleyn), turbulence (turbtran+turbdiff).
  • For the moment microphysics and radiation are in separate block loop. The turbulence scheme is copying 3d fields (i.e turbdiff(t(:,je,:) …)

Next steps

All 3 parametrizations (microphysics + radiation + turbulence) in a common block loop

Performance analysis

OMP parallelization (?)

Longer term

All parametrization required for operational runs should be inside the block loop and in 2 dimensional form

outline1
Outline
  • Physics with 2d data structure
  • Porting the physical parametrization to GPU using directives
  • Running COSMO on an hybrid GPU-CPU system
computing on graphical processing units gpus
Computing on Graphical Processing Units (GPUs)

Benefit from the highly parallel architecture of GPUs

Higher peak performance at lower cost / power consumption.

High memory bandwidth

execution model
Execution model

Data

Transfer

Parallel threads

Host

(CPU)

Device(GPU)

Copy data from CPU to GPU(CPU and GPU memory are separate)

Load specific GPU program (Kernel)

Execution: Same kernel is executed by all threads, SIMD parallelism (Single instruction, multiple data)

Copy back data from GPU to CPU

Sequential

Kernel

Sequential

the directive approach an example

3 different kernels

Array “a” remains on the GPU between the different kernel calls

The directive approach, an example

note : PGI directives

!$acc data region local(a,b)

!$acc update device(b)

!initialization

!$acc region do kerneldo i=1,Ndo k=1,nlev a(i,k)=0.0D0end do end do

!$acc end region

! first layer

!$acc region

do i=1,N a(i,1)=0.1D0end do

!$acc end region

! vertical computation

!$acc region do kernel

do i=1,N do k=2,nlev a(i,k)=0.95D0*a(i,k-1)+exp(-2*a(i,k)**2)*b(i,k) end do end do

!$acc end region

!$acc update host(a)

!$acc end data region

!$acc data region local(a,b)

!$acc update device(b)

!initialization

!$acc region

do k=1,nlevdo i=1,N a(i,k)=0.0D0end do end do

!$acc end region

! first layer

!$acc region

do i=1,N a(i,1)=0.1D0end do

!$acc end region

! vertical computation

!$acc region

do k=2,nlevdo i=1,N a(i,k)=0.95D0*a(i,k-1)+exp(-2*a(i,k)**2)*b(i,k) end do end do

!$acc end region

!$acc update host(a)

!$acc end data region

Loop reordering

N=1000, nlev=60: t= 555 μs t= 225 μs

physical parametrizations on gpu using directives
Physical parametrizations on GPU using directives
  • Physical parametrizations are tested using standalone code.
  • Currently ported parametrizations:
    • PGI : microphysics (hydci_pp), radiation (fesft), turbulence (only turbdiff yet)
    • OMP– acc (Cray) : microphysics, radiation
    • GPU optimizaiton: loop reordering, replacement of arrays with scalars
    • Note: hydci_pp, fesft and turbdiff subroutines represents respectively 6.7%, 8% and 7.3% of the total execution time of a typical cosmo-2 run.
  • Current version of OMP-acc are a subset of PGI directives and it is possible to write PGI code so that there is almost a one to one translation to omp-acc.
  • First investigation show similar performance between the two compilers, but would need further analysis
results fermi card using pgi directives
Results, Fermi card using PGI directives

Test domain: nx x ny x nz = 80 x 60 x 60

Peak performance of a Fermi card for double precision is 515 GFlop/s, i.e. we are getting respectively 5%, 4.5% and 2.5% peak performance for the microphysics, radiation and turbulence schemes

Theoretical bandwith is 140 GB/s, but maximum achievable is around 110 GB/s

results comparison with cpu
Results: Comparison with CPU

Parallel CPU code run on 12 cores AMD magny-cours CPU – however there are no mpi-communications in these standalone test codes.

Note: Expected performance would be between 3x and 5x and depending whether the problem is compute or memory bandwith bound.

Overhead of data transfer for microphysics and turbulence is very large.

comments on the observed performance
Comments on the observed performance
  • The microphysics has the largest compute intensity (with respect to memory access) and as such is more suited for the GPU.
  • The lower speed up observed for the radiation is quite relative, and essentially comes from the fact that it is very well optimized and is vectorized on the CPU (~9% Peak performance)
  • The turbulence scheme requires more memory access.

Next steps

Port turbtran subroutine with pgi + additional test and optimizations (october 2011)

Further investigation of radiation and turbulence schemes with Cray directives (november 2011)

GPU version of microphysics + radiation + turbulence inside COSMO (november-december 2011)

outline2
Outline
  • Physics with 2d data structure
  • Porting the physical parametrization to GPU using directives
  • Running COSMO on an hybrid GPU-CPU system
possible future implementations in cosmo
Possible future implementations in COSMO

Phys. parametrization

Dynamic

Microphysics

Turbulence

Radiation

I/O

GPU

GPU

GPU

GPU

Phys. parametrization

Dynamic

Microphysics

Turbulence

Radiation

I/O

GPU

Data movement for each routine

“Full GPU” : Data remain on device, only send to CPU for I/O and communication

Directives

C++ - CUDA

running cosmo 2 on hybrid system

GPUs

Running COSMO-2 on Hybrid-system

Multicores Processor

One (or more) multicores CPU

Domain decomposition

One GPU per subdomain.

summary
Summary

Porting of the microphysics, radiation and turbulence scheme on GPU was successfully carried out using a directive based approach

Comparing with a 12 cores CPU, a speed up between 2.4x and 6.5x was observed using one Fermi GPU card

These results are within the expected values considering hardware properties

The large overhead of data transfer shows that the “full GPU” approach (i.e. data remains on the GPU, all computation on the device) is the prefered approach for COSMO

comparison between pgi and omp acc
Comparison between PGI and OMP-acc

!$acc data region local(a)

!time loop

do itime=1,nt

!initialization

!$acc region

do k=1,nlevdo i=1,N a(i,k)=0.0D0end do end do

!$acc end region

! first layer

!$acc region do kernel

do i=1,N a(i,1)=0.1D0end do

!$acc end region

! vertical computation

!$acc region do kernel

do i=1,Ndo k=2,nlev a(i,k)=0.95D0*a(i,k-1)+exp(-2*a(i,k)**2)*a(i,k)end do end do

!$acc end region

end do ! end time loop

!$acc update host(a)

!$acc end data region

!$omp acc_data acc_shared(a)

!time loop

do itime=1,nt

!initialization

!$omp acc_region_loop

do k=1,nlevdo i=1,N a(i,k)=0.0D0end do end do

!$omp end acc_region loop

! first layer

!$omp acc_region_loop

do i=1,N a(i,1)=0.1D0end do

!$omp end acc_region_loop

! vertical computation

!$omp acc_region_loop kernel

do i=1,Ndo k=2,nlev a(i,k)=0.95D0*a(i,k-1)+exp(-2*a(i,k)**2)*a(i,k)end do end do

!$omp end acc_region_loop

end do ! end time loop

!$omp acc_update host(a)

!$omp end acc_data

slide19

Craypat infos

MAIN_ / mo_gscp_dwd_hydci_pp_ _ (x10)

------------------------------------------------------------------------

User time (approx) 2.999 secs 7197500711 cycles

System to D1 refill 2.434M/sec 7300271 lines

System to D1 bandwidth 148.576MB/sec 467217344 bytes

D2 to D1 bandwidth 1025.770MB/sec 3225672832 bytes

L2 to System BW per core 140.940MB/sec 443203504 bytes

HW FP Ops / User time 435.162M/sec 1308546592 ops 4.5%peak(DP)

MAIN_ / src_radiation_fesft_ (x1)

------------------------------------------------------------------------

User time (approx) 7.226 secs 17342858074 cycles 100.0%Time

System to D1 refill 11.380M/sec 82232710 lines

System to D1 bandwidth 694.569MB/sec 5262893440 bytes

D2 to D1 bandwidth 1162.252MB/sec 8806624128 bytes

L2 to System BW per core 645.679MB/sec 4892446080 bytes

HW FP Ops / User time 893.252M/sec 6511701846 ops 9.3%peak(DP)

MAIN_ / turbulence_diff_ref_turbdiff_ (x10)

------------------------------------------------------------------------

User time (approx) 4.397 secs 10551890928 cycles 100.0%Time

System to D1 refill 15.757M/sec 69278266 lines

System to D1 bandwidth 961.741MB/sec 4433809024 bytes

D2 to D1 bandwidth 485.462MB/sec 2238073856 bytes

L2 to System BW per core 982.474MB/sec 4529394160 bytes

HW FP Ops / User time 326.405M/sec 1452061875 ops 3.4%peak(DP)