Utilizing Intel Xeon Phi Server

We have a new server that is equipped with 2 Xeon Phi 5110 coprocessors (MIC = Many Integrated Cores).

You are welcome to try out this node and see whether this can accelerate your code.

Information about the Intel Xeon Phi coprocessor is given here:

http://www.intel.com/content/www/us/en/high-performance-computing/high-performance-xeon-phi-coprocessor-brief.html

In essence, it can accelerate your code by running on high-number-of threads coprocessors, similar to GPU, but now running on similar Intel structure. And instead of having to write a code separately for the GPU, you can use your code, insert some MIC directives within the code, and then you can reap the benefit of getting the code accelerated.

Here is a book about Xeon Phi programming that would possibly help you get started:

Intel Xeon Phi Coprocessor High Performance Programming, Jeffers and Reinders, 2012

Another difference from the GPU is the ability for the users to ssh directly to the MICs (mic0 and mic1) and run very basic Busybox Linux commands, while in this mode.

EXAMPLES

Request the node using qsub:

qsub -I -q phi -l nodes=1:ppn=12:mics=2

There is a directory containing the code examples that you can copy, compile and run for your testing.

Copy the codes from /usr/local/doc/phi/uncompiled/* to your home directory (e.g. /home/hxd58/test/phi/.)

Compiling the codes (for example if we do this in the directory .../uncompiled/Ch2):

cp -r /usr/local/doc/phi/uncompiled/* ~/test/phi/.

cd ~/test/phi/uncompiled/Ch2

source /usr/local/intel/2013/composer_xe_2013/bin/compilervars.sh intel64

make all

Please change the source location to compilervars.csh if you are using C-shell.

Options:

• 1. Running directly on the MIC (Xeon Phi)

     1. ssh mic0

     2. cd ~/test/phi/uncompiled/Ch2

     3. ./helloflops1_xphi

     4. Initializing

     5. Starting Compute

     6. GFlops =     25.600, Secs =      1.530, GFlops per sec =     16.730

  2. Offload the code from the node and let it run on the MIC

• • 1. module load micset

    2. cd ~/test/phi/uncompiled/Ch2

    3. ./helloflops3o_xeon

    4. Initializing

    5. Starting Compute on 236 threads

    6. GFlops =   6041.600, Secs =      3.124, GFlops per sec =   1933.704

You should be able to compile and runs for codes in Ch3 and Ch4, which runs directly on the MIC (Xeon Phi).

Results of these test codes are presented in this Google Spreadsheet:

https://docs.google.com/a/case.edu/spreadsheet/ccc?key=0Ai0GTaYFOIhddGNRdmYtVy10Z05ObTJYOG5ZUFpEV1E&usp=drive_web#gid=0