SHOC Alpha v1.1.4a-mic

Alpha Status

This release is an alpha. It is expected to have bugs and has been tested on a limited number of platforms compared to the main distribution. The supported platform is CentOS 6.2 \ Intel Compiler v13.0.1

Please contact shoc-help@elist.ornl.gov with any questions concerning this release.

Please see the release notes below for other limitations and known issues.

Building the Benchmarks

Type make inside this directory. This assumes the intel toolchain (a recent version of icc) is in your current path. Your paths should also include MKL, which is used in some of the benchmarks.

Running the Benchmarks

How to run all the benchmarks:

1) Use the provided run script in /bin

How to run an individual benchmark:

2) Set the appropriate environment variables (see run scripts)

For example, to run MD workload

    $ export MIC_ENV_PREFIX=MIC
    $ export MIC_USE_2MB_BUFFERS=32K
    $ export MIC_KMP_AFFINITY=granularity=fine,balanced
    $ export MIC_BUFFERSIZE=128M
    $ export nc=61
    $ export MIC_OMP_NUM_THREADS=4*$nc-4
    $ ./MD -s 4

For another example, to run FFT workload

    $ export MIC_ENV_PREFIX=MIC 
    $ export MIC_USE_2MB_BUFFERS=32K 
    $ export MIC_KMP_AFFINITY=granularity=fine,balanced 
    $ export MIC_BUFFERSIZE=128M 
    $ export MIC_MKL_DYNAMIC=false
    $ nc=61
    $ export MIC_OMP_NUM_THREADS=4*$nc-4
    $ ./FFT -s 4

The value for -s is a number between 1 and four that corresponds to problem size from smallest to largest (consistent with the CUDA and OpenCL versions).

Release Notes

• This is a preview release and does not support MPI or more than one MIC card per node.
• The current benchmarks have not yet been integrated into the main SHOC infrastructure. As such, a simplified runscript has been provided in the /bin directory.
• MIC versions of the benchmarks have can have (much) longer build\run times than CUDA/OCL. In particular, the compile time for S3D and the runtime for DeviceMemory are quite long.

Known Issues

• DeviceMemory is inconsistent with CUDA/OpenCL. Its tests are structured to be specific for MIC hardware.
• Current release uses gettimeofday() exclusively. In most cases this shouldn't matter.
• Sort does not currently complete all radix passes.