Skip to content
/ NPB-GPU Public

NAS Parallel Benchmarks for evaluating GPU and APIs

License

Notifications You must be signed in to change notification settings

GMAP/NPB-GPU

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

36 Commits
 
 
 
 
 
 
 
 

Repository files navigation

NAS Parallel Benchmarks for GPU

This is a repository aimed at providing GPU parallel codes with different parallel APIs for the NAS Parallel Benchmarks (NPB) from a C/C++ version (NPB-CPP). You can also contribute with this project, writing issues and pull requests. 😄

🔉News: CUDA versions for pseudo-applications added and IS improved. 📅11/Feb/2021

🔉News: Parametrization support for configuring number of threads per block and CUDA parallelism optimizations. 📅25/Jul/2021

🔉News: Paper published in the journal Software: Practice and Experience (SPE). 📅29/Nov/2021

🔉News: A new GPU parallel implementation is now available using the GSParLib API. 📅15/Aug/2024

How to cite our work 👍

DOI - Araujo, G.; Griebler, D.; Rockenbach, D. A.; Danelutto, M.; Fernandes, L. G.; NAS Parallel Benchmarks with CUDA and beyond, Software: Practice and Experience (SPE), 2021.

DOI - Araujo, G.; Griebler, D.; Danelutto, M.; Fernandes, L. G.; Efficient NAS Benchmark Kernels with CUDA. 28th Euromicro International Conference on Parallel, Distributed and Network-based Processing (PDP), Västerås, 2020.

The NPB with CUDA

The parallel CUDA version was implemented from the serial version of NPB-CPP.

==================================================================

NAS Parallel Benchmarks code contributors with CUDA are:

Dalvan Griebler: [email protected]

Gabriell Araujo: [email protected]

==================================================================

Each directory is independent and contains its own implemented version:

Five kernels

  • IS - Integer Sort
  • EP - Embarrassingly Parallel
  • CG - Conjugate Gradient
  • MG - Multi-Grid
  • FT - discrete 3D fast Fourier Transform

Three pseudo-application

  • SP - Scalar Penta-diagonal solver
  • BT - Block Tri-diagonal solver
  • LU - Lower-Upper Gauss-Seidel solver

Software Requiriments

Warning: our tests were made with GCC and CUDA

How to Compile

  1. Update the make.def file with the compute capability of the GPU you want to use to compile and run the NPB-GPU:

    1. Check the number of available NVIDIA GPUs by executing the following command:

      nvidia-smi --query-gpu=index --format=csv,noheader,nounits | wc -l
      
    2. Find the compute capability of the GPU you want to use by executing the following command (replace GPU_ID with the actual GPU ID):

      nvidia-smi -i GPU_ID --query-gpu=compute_cap --format=csv,noheader,nounits
      
    3. Update the make.def file by replacing 61 in the following line with the value of GPU compute compute capability you obtained:

      COMPUTE_CAPABILITY = -gencode arch=compute_61,code=sm_61
      
  2. Go inside the CUDA directory and execute:

    make _BENCHMARK CLASS=_VERSION
    

    _BENCHMARKs are:

     CG, EP, FT, IS, MG, BT, LU, and SP 
    

    _VERSIONs are:

     + Class S: small for quick test purposes
    
     + Class W: workstation size (a 90's workstation; now likely too small)
    
     + Classes A, B, C: standard test problems; ~4X size increase going from one class to the next
    
     + Classes D, E, F: large test problems; ~16X size increase from each of the previous Classes
    

    Command example:

    make ep CLASS=B
    

Activating the additional timers

NPB-GPU has additional timers for profiling purpose. To activate these timers, create a dummy file 'timer.flag' in the main directory of the NPB version (e.g. CUDA/timer.flag).

Configuring the number of threads per block

NPB-GPU allows configuring the number of threads per block of each GPU kernel in the benchmarks. The user can specify the number of threads per block by editing the file gpu.config in the directory /config/. If no file is specified, all GPU kernels are executed using the warp size of the GPU as the number of threads per block.

Syntax of the gpu.config file:

<benchmark-name>_THREADS_PER_BLOCK_<gpu-kernel-name> = <interger-value>

Configuring CG benchmark as example:

CG_THREADS_PER_BLOCK_ON_KERNEL_ONE = 32
CG_THREADS_PER_BLOCK_ON_KERNEL_TWO = 128
CG_THREADS_PER_BLOCK_ON_KERNEL_THREE = 64
CG_THREADS_PER_BLOCK_ON_KERNEL_FOUR = 256
CG_THREADS_PER_BLOCK_ON_KERNEL_FIVE = 32
CG_THREADS_PER_BLOCK_ON_KERNEL_SIX = 64
CG_THREADS_PER_BLOCK_ON_KERNEL_SEVEN = 128
CG_THREADS_PER_BLOCK_ON_KERNEL_EIGHT = 64
CG_THREADS_PER_BLOCK_ON_KERNEL_NINE = 512
CG_THREADS_PER_BLOCK_ON_KERNEL_TEN = 512
CG_THREADS_PER_BLOCK_ON_KERNEL_ELEVEN = 1024

The NPB-GPU also allows changing the GPU device by providing the following syntax in the gpu.config file:

GPU_DEVICE = <interger-value>