GPU-STREAM

Measure memory transfer rates to/from global device memory on GPUs. This benchmark is similar in spirit, and based on, the STREAM benchmark [1] for CPUs.

Unlike other GPU memory bandwidth benchmarks this does not include the PCIe transfer time.

Usage

Build the OpenCL and CUDA binaries with make (CUDA version requires CUDA >= v6.5)

Run the OpenCL version with ./gpu-stream-ocl and the CUDA version with ./gpu-stream-cuda

Automatic variation of array size

I added a bash script that automatically re-runs ./gpu-stream-ocl with different array size and prints out results in columns, useful for plotting figures.

./run-ocl.sh

# Benchmark GPU-STREAM running on  Tesla C2070
# Precision: double. Range: [204800 .. 102400000] step 409600
# For more details see https://github.com/UoB-HPC/GPU-STREAM
#  ArrayElements    ArraySize(MB)      Copy(MBytes/s)    Mul(MBytes/s)      Add(MBytes/s)      Triad(MBytes/s)
    204800            1.56             80753.117          79927.800          83088.782          83981.752          
    614400            4.68             93275.517          93897.395          94380.296          93960.518          
    1024000            7.81             98982.027          98607.910          97231.728          97690.504          
    1433600            10.93             99431.697          100773.675          99081.649          98843.119          
    1843200            14.06             100561.265          101067.177          99730.141          99793.135          
    ...         
    4300800            32.81             101984.021          101605.451          100626.557          100612.825          
    4710400            35.93             102605.629          103710.187          100675.208          100587.601          
    5120000            39.06             101817.603          102083.414          101225.864          101156.117 

Android

Assuming you have a recent Android NDK available, you can use the toolchain that it provides to build GPU-STREAM. You should first use the NDK to generate a standalone toolchain:

# Select a directory to install the toolchain to
ANDROID_NATIVE_TOOLCHAIN=/path/to/toolchain

${NDK}/build/tools/make-standalone-toolchain.sh \
  --platform=android-14 \
  --toolchain=arm-linux-androideabi-4.8 \
  --install-dir=${ANDROID_NATIVE_TOOLCHAIN}

Make sure that the OpenCL headers and library (libOpenCL.so) are available in ${ANDROID_NATIVE_TOOLCHAIN}/sysroot/usr/.

You should then be able to build GPU-STREAM:

make CXX=${ANDROID_NATIVE_TOOLCHAIN}/bin/arm-linux-androideabi-g++

Copy the executable and OpenCL kernels to the device:

adb push gpu-stream-ocl /data/local/tmp
adb push ocl-stream-kernels.cl /data/local/tmp

Run GPU-STREAM from an adb shell:

adb shell
cd /data/local/tmp

# Use float if device doesn't support double, and reduce array size
./gpu-stream-ocl --float -n 6 -s 10000000

Results

Sample results can be found in the results subdirectory. If you would like to submit updated results, please submit a Pull Request.

[1]: McCalpin, John D., 1995: "Memory Bandwidth and Machine Balance in Current High Performance Computers", IEEE Computer Society Technical Committee on Computer Architecture (TCCA) Newsletter, December 1995.