GPU Devel Nodes

GPU Development Cluster (ARC)

Installed	April 2011
Operating System	Linux Centos 6.4
Number of Nodes	8 (64 cpu cores, 7168 gpu cores)
Interconnect	DDR Infiniband
Ram/Node	48 Gb
Cores/Node	8 with 2xGPUs
Login/Devel Node	arc01 (from `login.scinet`)
Vendor Compilers	nvcc,pgcc,icc,gcc
Queue Submission	Torque

The GPU cluster, part of the Accelerator Research Cluster, consists of 8 x86_64 nodes each with two quad core Intel Xeon X5550 2.67GHz CPUs with 48GB of RAM per node. Each node has two NVIDIA Tesla M2070 GPUs with CUDA Capability 2.0 (Fermi) each with 448 CUDA Cores @ 1.15GHz and 6 GB of RAM. The nodes are interconnected with DDR Infiniband for MPI communications and disk I/O to the SciNet GPFS filesystems. In total this cluster contains 64 x86_64 cores with 384 GB of system RAM and 16 GPUs with 96 GB GPU RAM total.

Note that SciNet has a mailing lists for people interested in GPGPU computing. To receive information on courses, workshop and other GPGPU related events, sign up at https://support.scinet.utoronto.ca/mailman/listinfo/scinet-gpgpu.

Nodes

Login

First login via ssh with your scinet account at login.scinet.utoronto.ca, and from there you can proceed to arc01 which is the GPU development node.

Access to these machines is currently controlled. Please email support@scinet.utoronto.ca for access.

Devel

As mentioned arc01 is the head/develop node for interactive use. This node is for compiling, short testing, and submitting batch jobs to the compute nodes. It is a shared resource so treat it accordingly and use the queue and compute nodes for long are large computations.

ARC Experimental (ARCX) Xeon Phi/ Tesla K20

A separate devel node, arcX, with a single Intel Xeon Phi and a NVIDIA Tesla K20 is also available for testing these newer technologies. For full details see the Xeon Phi / Tesla K20 wiki page.

Compute

To access the other 7 compute nodes with GPU's you need to use the queue, similar to the standard GPC compute nodes. Currently the nodes are scheduled by complete node, 8 cores and 2 GPUs, a minimum of 15 minutes and a maximum walltime of 48 hours.

For an interactive job use

qsub -l nodes=1:ppn=8:gpus=2,walltime=48:00:00 -q arc -I

or for a batch job use

qsub script.sh

where script.sh is <source lang="bash">

!/bin/bash
Torque submission script for SciNet ARC
PBS -l nodes=2:ppn=8:gpus=2,walltime=1:00:00
PBS -N GPUtest
PBS -q arc

cd $PBS_O_WORKDIR

module load intel openmpi cuda

EXECUTION COMMAND; -np = nodes*ppn

mpirun -np 16 ./a.out </source>

To check running jobs on the gpu nodes only use

showq -w class=arc

Important note:

A bug in the torque scheduler currently sets the environment variable CUDA_VISIBLE_DEVICES to an incorrect value. Loading any one of the cuda modules will correct this, so be to do this in your job script or in your interactive jobs.

Software

The same software installed on the GPC is available on ARC using the same modules framework. See here for full details.

Programming Frameworks

Currently there are four programming frameworks to use: NVIDIA's CUDA framework, PGI's CUDA Fortran, PGI's implementation of OpenACC, or OpenCL.

NVIDIA toolkit

CUDA

The current installed CUDA Toolkits are 3.2, 4.0, 4.1 (default), 4.2, 5.0, 5.5 , and 6.0. To use 6.0 just add the following module

module load cuda/6.0

Note that to use the full 6GB or memory per GPU, CUDA 3.2 or newer must be used.

The CUDA driver is installed locally, however the CUDA Toolkits are installed in.

/scinet/arc/cuda-$VERSION/

The environment variable $SCINET_CUDA_INSTALL is set when a cuda module is loaded and it points to the install location. This is useful when setting up makefiles and if you use the NVIDIA_SDK build evironment, modify the NVIDIA_SDK/C/common/common.mk file accordingly.

CUDA_INSTALL_PATH = $SCINET_CUDA_INSTALL

The Nvidia cuda compiler (which uses gcc/4.4.6 by default for CUDA < 4.1, while cuda/4.2 uses gcc/4.6.1), is called nvcc,

You'll have to let the cuda compiler know about the capabilities of the Fermi graphics card by supplying the flag

-arch=sm_13

-arch=sm_20

NVIDIA Toolkit

For cuda versions 4.0, 4.1, and 4.2, the CUDA SDK can be copied into your home directory from:

/scinet/arc/src/gpucomputingsdk_4.0.13_linux.run
/scinet/arc/src/gpucomputingsdk_4.1.28_linux.run
/scinet/arc/src/gpucomputingsdk_4.2.9_linux.run

respectively.

However, for cuda 5.0 and 5.5, the sdk code samples can be copied from the directory

$SCINET_CUDA_INSTALL/samples/

NOTE: Not all of the CUDA and OpenCL examples will compile as many require OpenGL graphic libraries not installed on the nodes.

OpenCL

As of 3.0, OpenCL 1.1 is included in the CUDA Toolkit so loading the CUDA module is all that is required.

PGI compilers

As of July 2012, The PGI suite of compilers is installed on the ARC. These can be accessed by

$  module load gcc/4.6.1 pgi/12.6

(if you use the older pgi/12.5, gcc/4.4.6 is a requirement, and is used, for instance, in the CUDA parts of the PGI compilers). These compilers use their own cuda installation, so you do not need to load an additional cuda module. By default, they use a cuda 4.1 installation, but you can request cuda 4.2 as well using the -Mcuda=4.2 option.

The compilation commands are pgcc, pgcpp and pgfortran for c, c++ and fortran, respectively. As usual, we advice to compiler with optimization using the flags

-O4 -fastsse

The compilers will then optimize for the specific machine that you are compiling on.

The PGI compilers support OpenMP as well through the compile and link flags

-mp

CUDA Fortran

The PGI fortran compiler (pgfortran, also pgf77 and pgf90) understands CUDA extensions to fortran. This compiler will automatically understand these extension for source files with the file extension .cuf Otherwise, you have to specify

-Mcuda=4.1

OpenACC

OpenACC is a compiler-directive approach to GPGPU programming. The PGI compilers (c, c++ and fortran) have a partial implementation of this open specification. To switch this on, use the options

-acc -ta=nvidia -Mcuda=4.1

Other compilers

gcc,g++,gfortran - GNU compiler (nvcc need to have either gcc-4.4 or gcc-4.6 module loaded to work correctly)
icc,icpc,ifort - Intel compiler

Debuggers

ddt - Allinea's graphical DDT debugger, in the ddt module. The most recent version, ddt 4.0 supports cuda 4.0, 4.1, 4.2 and 5.0.
cuda-gdb - Nvidia text based gdb variant, part of the cuda module.

Note that to debug both host and cuda device code, you have to give the

-g -G

pair of flags to nvcc.

MPI

The GPC MPI packages can be used on this system. See the GPC section on MPI for more details.

While these mpi packages should work with the PGI compilers as well, this has not been tested and standard wrappers like mpif90 may not work.

Alternatively, for mpi compilations with the PGI compilers, you can load the mpich1 mpi implementation with

module load mpich1/pgi

after which you can use the option

-Mmpi

or the wrapper scripts mpicc, mpiCC and mpif90, as well as mpirun.

Driver Version

The current NVIDIA driver version installed is 295.41.

Documentation

CUDA
- google "CUDA"

OpenCL
- see above

Further Info

User Codes

Please discuss and put any relevant information/problems/best practices you have encountered when using/developing for CUDA and/or OpenCL

Contents