mgherghi/intel-xeon-phi

Examples for running software on Intel Xeon Phi Co-Processors

Intel Xeon Phi Co-Processors on ACCRE

• The Intel Xeon Phi co-processor (aka, Many Integrated Core, or MIC card) is a device that can be used to improve the performance of compute-intensive applications. It can work independently or in tandem with a CPU (aka, host) to reduce execution time of certain programs, especially those that rely on Intel's MKL (Math Kernel Library) or multi-threading (especially via OpenMP) to process data in parallel across multiple CPU cores.

• The biggest advantage of the Phis is the ability to achieve higher performance with minimal changes to your code. In particular, programs that perform large matrix and/or vector processing using Intel's MKL or multi-threading are execellent candidates for Phi execution. Intel's "automatic offloading" technology is enabled on the cluster for at least three widely used software environments (Matlab, Python, and R), where a user only needs to add a few lines to a SLURM script to achieve dynamic offloading to the Phi. No re-building/compiling of the software is needed.

• This repository contains examples and instructions for running programs on the ACCRE cluster using Intel Xeon Phis. Click on the directories above from Github to view these simple examples.

• If you are a Vanderbilt University researcher, you can register for a Intel Xeon Phi training course here: http://www.accre.vanderbilt.edu/?page_id=3019

Available Phi-Supported Software on ACCRE Cluster

The following software packages support automatic offloading to the Phi, whereby large matrix and vector operations can be dynamically offloaded to the Phi simply by adding a few lines of code to your SLURM script. LAMMPS does not technically support automatic offloading but it does include Phi support for certain algorithms.

• Matlab (r2014a and later)

setpkgs -a matlab

• R (3.1.1 and 3.2.0)

setpkgs -a R_3.1.1_intel
setpkgs -a R_3.2.0

• Python (2.7.8) / NumPy

setpkgs -a python2.7.8_intel14

• LAMMPS (mid-July 2015 version)

setpkgs -a lammps_mic

Examples for each of these are included in this repo. Each of these packages can take advantage of both Phis residing on ACCRE Phi nodes and will run programs across both the host (normal CPU) and the Phis simultaneously without much intervention from the user.

In addition to these automatic offloading examples, there are also a few C programs demonstrating how to run programs either natively on a MIC card or by explicitly offloading certain tasks from the host to a MIC card. These examples also contain Makefiles to aid in building the executables for these examples. See the directories labeled:

• Native
• Offload

Getting Access

Add the following line to your SLURM script to submit to the MIC queue:

#SBATCH --partition=mic

If you get an error like this one:

sbatch: error: Batch job submission failed: Invalid account or account/partition combination specified

Then open a ticket with us (http://www.accre.vanderbilt.edu/?page_id=369) requesting to be granted access to the Intel Xeon Phi nodes.

Cluster Policies

ACCRE currently has four Phi nodes in production. Each node contains two Phi co-processors. Until usage on the Phi nodes gets high enough, a single job will be allocated an entire node, which consists of:

• 2 Intel Xeon Phi Co-Processors (each with 61 cores, 4 hardware threads/core, 1.24 GHz)
• 2 Intel Xeon E5-2670 CPUs (each with 8 cores, 2 hardware threads/core, 2.60 GHz)
• 132 GB system RAM
• 15.8 GB RAM per Phi card

By default a job will have access to both Phis, all 8 CPU cores, and 16 GB of MIC RAM. To request more system memory simply add a #SBATCH --mem= directive to your SLURM script. Some memory is reserved for the OS and filesystem so you cannot request any more than 123 GB.