Difference between revisions of "SOSCIP GPU"

From oldwiki.scinet.utoronto.ca
Jump to navigation Jump to search
 
(71 intermediate revisions by 5 users not shown)
Line 1: Line 1:
 +
__NOTOC__
 +
 +
{| style="border-spacing: 8px; width:100%"
 +
| valign="top" style="cellpadding:1em; padding:1em; border:2px solid; background-color:#f6f674; border-radius:5px"|
 +
'''WARNING: SciNet is in the process of replacing this wiki with a new documentation site. For current information, please go to [https://docs.scinet.utoronto.ca https://docs.scinet.utoronto.ca]'''
 +
|}
 +
 
{{Infobox Computer
 
{{Infobox Computer
 
|image=[[Image:S882lc.png|center|300px|thumb]]
 
|image=[[Image:S882lc.png|center|300px|thumb]]
Line 11: Line 18:
 
|vendorcompilers=xlc/xlf, nvcc
 
|vendorcompilers=xlc/xlf, nvcc
 
}}
 
}}
 +
 +
== New Documentation Site ==
 +
Please visit the new documentation site: [https://docs.scinet.utoronto.ca/index.php/SOSCIP_GPU https://docs.scinet.utoronto.ca/index.php/SOSCIP_GPU] for updated information.
  
 
== SOSCIP ==
 
== SOSCIP ==
Line 20: Line 30:
 
Please use [mailto:soscip-support@scinet.utoronto.ca <soscip-support@scinet.utoronto.ca>] for SOSCIP GPU specific inquiries.
 
Please use [mailto:soscip-support@scinet.utoronto.ca <soscip-support@scinet.utoronto.ca>] for SOSCIP GPU specific inquiries.
  
 +
 +
<!--
 
== Specifications==
 
== Specifications==
  
 
The SOSCIP GPU Cluster consists of  of 14 IBM Power 822LC "Minsky" Servers each with 2x10core 3.25GHz Power8 CPUs and 512GB Ram. Similar to Power 7, the Power 8 utilizes Simultaneous MultiThreading (SMT), but extends the design to 8 threads per core allowing the 20 physical cores to support up to 160 threads.  Each node has 4x NVIDIA Tesla P100 GPUs each with 16GB of RAM with CUDA Capability 6.0 (Pascal) connected using NVlink.
 
The SOSCIP GPU Cluster consists of  of 14 IBM Power 822LC "Minsky" Servers each with 2x10core 3.25GHz Power8 CPUs and 512GB Ram. Similar to Power 7, the Power 8 utilizes Simultaneous MultiThreading (SMT), but extends the design to 8 threads per core allowing the 20 physical cores to support up to 160 threads.  Each node has 4x NVIDIA Tesla P100 GPUs each with 16GB of RAM with CUDA Capability 6.0 (Pascal) connected using NVlink.
  
−
== Compile/Devel/Test ==
+
== Access and Login ==
  
−
Access is provided through the BGQ login node, '''<tt> bgqdev.scinet.utoronto.ca </tt>''' using ssh, and from there you can proceed to the GPU development node '''<tt>sgc01-ib0</tt>'''.
+
In order to obtain access to the system, you must request access to the SOSCIP GPU Platform. Instructions will have been sent to your sponsoring faculty member via E-mail at the beginning of your SOSCIP project.
 +
 
 +
Access to the SOSCIP GPU Platform is provided through the BGQ login node, '''<tt> bgqdev.scinet.utoronto.ca </tt>''' using ssh, and from there you can proceed to the GPU development node '''<tt>sgc01-ib0</tt>''' via ssh. Your user name and password is the same as it is for SciNet systems.
  
 
== Filesystem ==
 
== Filesystem ==
Line 55: Line 69:
 
</pre>
 
</pre>
  
−
You can queury job information using
+
More information about the <tt>sbatch</tt> command is found [https://slurm.schedmd.com/sbatch.html here].
 +
 
 +
 
 +
You can query job information using
  
 
<pre>
 
<pre>
 
squeue
 
squeue
 
</pre>
 
</pre>
 +
 +
To see only your own jobs, run
 +
 +
<pre>
 +
squeue -u <userid>
 +
</pre>
 +
 +
Once your job is running, SLURM creates a file usually named <tt>slurm<jobid>.out</tt> in the directory from where you issued the <tt>sbatch</tt> command. This contains the console output from your job. You can monitor the output of your job by using the <tt>tail -f <file></tt> command.
 +
  
 
To cancel a job use
 
To cancel a job use
Line 78: Line 104:
 
salloc --gres=gpu:4
 
salloc --gres=gpu:4
 
</pre>
 
</pre>
 +
 +
After executing this command, you may have to wait in the queue until a system is available.
 +
 +
More information about the <tt>salloc</tt> command is [https://slurm.schedmd.com/salloc.html here].
  
 
=== Automatic Re-submission and Job Dependencies ===
 
=== Automatic Re-submission and Job Dependencies ===
Line 114: Line 144:
  
 
</pre>
 
</pre>
 +
===Packing single-GPU jobs within one SLURM job submission===
 +
Jobs are scheduled by node (4 GPUs) on SOSCIP GPU cluster. If user's code/program cannot utilize all 4 GPUs, user can use GNU Parallel tool to pack 4 or more single-GPU jobs into one SLURM job. Below is an example of submitting 4 single-GPU python codes within one job:  (When using GNU parallel for a publication please cite as per '''''parallel --citation''''')
 +
<pre>
 +
#!/bin/bash
 +
#SBATCH --nodes=1
 +
#SBATCH --ntasks=20  # MPI tasks (needed for srun)
 +
#SBATCH --time=00:10:00  # H:M:S
 +
#SBATCH --gres=gpu:4    # Ask for 4 GPUs per node
  
−
== Software ==
+
module load gnu-parallel/20180422
 +
cd $SLURM_SUBMIT_DIR
  
−
==== GNU Compilers ====
+
parallel -a jobname-params.input --colsep ' ' -j 4 'CUDA_VISIBLE_DEVICES=$(( {%} - 1 )) numactl -N $(( ({%} -1) / 2 )) python {1} {2} {3} &> jobname-{#}.out'
 +
</pre>
 +
The jobname-params.input file contains:
 +
<pre>
 +
code-1.py --param1=a --param2=b
 +
code-2.py --param1=c --param2=d
 +
code-3.py --param1=e --param2=f
 +
code-4.py --param1=g --param2=h
 +
</pre>
 +
*In the above example, GNU Parallel tool will read '''jobname-params.input''' file and separate parameters. Each row in the input file has to contain exact 3 parameters to '''python'''. code-N.py is also considered as a parameter. User can change parameter number in the '''parallel''' command ({1} {2} {3}...).
 +
*'''"-j 4"''' flag limits the max number of jobs to be 4. User can have more rows in the input file, but GNU Parallel tool only executes maximum of 4 at the same time.
 +
*'''"CUDA_VISIBLE_DEVICES=$(( {%} - 1 ))"''' will set one GPU for each job. '''"numactl -N $(( ({%} -1) / 2 ))"''' will bind 2 jobs on CPU socket 0, other 2 jobs on socket 1. {%} is job slot which will be translated to 1 or 2 or 3 or 4 in this case.
 +
*Outputs will be  jobname-1.out, jobname-2.out,jobname-3.out,jobname-4.out... {#} is job number which will be translated to the row number in the input file.
  
−
More recent versions of the GNU Compiler Collection (C/C++/Fortran) are provided in the IBM Advanced Toolchain with enhancements for the POWER8 CPU. To load the newer advance toolchain version use:
+
== Software Installed ==
  
 +
=== IBM PowerAI ===
 +
 +
The PowerAI platform contains popular open machine learning frameworks such as '''Caffe, TensorFlow, and Torch'''. Run the <tt>module avail</tt> command for a complete listing. More information is available at this link: https://developer.ibm.com/linuxonpower/deep-learning-powerai/releases/. Release 4.0 is currently installed.
 +
 +
===GNU Compilers ===
 +
 +
System default compiler is GCC/5.4.0. More recent versions of the GNU Compiler Collection (C/C++/Fortran) are provided in the IBM Advance Toolchain with enhancements for the POWER8 CPU. To load the newer advance toolchain version use:
 +
 +
Advance Toolchain V10.0
 
<pre>
 
<pre>
−
module load gcc/6.3.1
+
module load gcc/6.4.1
 
</pre>
 
</pre>
  
−
More information about the IBM Advanced Toolchain can be found here: [https://developer.ibm.com/linuxonpower/advance-toolchain/ https://developer.ibm.com/linuxonpower/advance-toolchain/]
+
Advance Toolchain V11.0
 +
<pre>
 +
module load gcc/7.3.1
 +
</pre>
  
−
==== IBM Compilers ====
+
More information about the IBM Advance Toolchain can be found here: [https://developer.ibm.com/linuxonpower/advance-toolchain/ https://developer.ibm.com/linuxonpower/advance-toolchain/]
  
−
To load the native IBM xlc/xlc++ compilers
+
=== IBM XL Compilers ===
 +
 
 +
To load the native IBM xlc/xlc++ and xlf (Fortran) compilers, run
  
 
<pre>
 
<pre>
Line 136: Line 201:
 
</pre>
 
</pre>
  
−
==== Driver Version ====
+
IBM XL Compilers are enabled for use with NVIDIA GPUs, including support for OpenMP 4.5 GPU offloading and integration with NVIDIA's nvcc command to compile host-side code for the POWER8 CPU.
  
−
The current NVIDIA driver version is 384.66
+
Information about the IBM XL Compilers can be found at the following links:
  
−
==== CUDA ====
+
[https://www.ibm.com/support/knowledgecenter/SSXVZZ_13.1.5/com.ibm.compilers.linux.doc/welcome.html IBM XL C/C++]
  
−
The current installed CUDA Tookit is 8.0
+
[https://www.ibm.com/support/knowledgecenter/SSAT4T_15.1.5/com.ibm.compilers.linux.doc/welcome.html IBM XL Fortran]
 +
 
 +
=== NVIDIA GPU Driver ===
 +
 
 +
The current NVIDIA driver version is 396.26
 +
 
 +
=== CUDA ===
 +
 
 +
The current installed CUDA Tookits is are version 8.0, 9.0 and 9.1.
  
 
<pre>
 
<pre>
 
module load cuda/8.0
 
module load cuda/8.0
 +
or
 +
module load cuda/9.0
 +
or
 +
module load cuda/9.1
 +
or
 +
module load cuda/9.2
 
</pre>
 
</pre>
 +
  
 
The CUDA driver is installed locally, however the CUDA Toolkit is installed in:
 
The CUDA driver is installed locally, however the CUDA Toolkit is installed in:
Line 152: Line 232:
 
<pre>
 
<pre>
 
/usr/local/cuda-8.0
 
/usr/local/cuda-8.0
 +
/usr/local/cuda-9.0
 +
/usr/local/cuda-9.1
 +
/usr/local/cuda-9.2
 
</pre>
 
</pre>
  
−
==== OpenMPI ====
+
Note that the <tt>/usr/local/cuda</tt> directory is linked to the <tt>/usr/local/cuda-9.2</tt> directory.
 +
 
 +
Documentation and API reference information for the CUDA Toolkit can be found here: [http://docs.nvidia.com/cuda/index.html http://docs.nvidia.com/cuda/index.html]
 +
 
 +
=== OpenMPI ===
  
 
Currently OpenMPI has been setup on the 14 nodes connected over EDR Infiniband.
 
Currently OpenMPI has been setup on the 14 nodes connected over EDR Infiniband.
Line 163: Line 250:
 
</pre>
 
</pre>
  
−
=== IBM PowerAI ===
+
== Other Software ==
 +
 
 +
Other software packages can be installed onto the SOSCIP GPU Platform. It is best to try installing new software in your own home directory, which will give you control of the software (e.g. exact version, configuration, installing sub-packages, etc.).
 +
 
 +
In the following subsections are instructions for installing several common software packages.
 +
 
 +
=== Anaconda (Python) ===
 +
 
 +
Anaconda is a popular distribution of the Python programming language. It contains several common Python libraries such as SciPy and NumPy as pre-built packages, which eases installation.
 +
 
 +
Anaconda can be downloaded from here: [https://www.anaconda.com/download/#linux https://www.anaconda.com/download/#linux]
 +
 
 +
NOTE: Be sure to download the '''Power8''' installer.
 +
 
 +
TIP: If you plan to use Tensorflow within Anaconda, download the Python 2.7 version of Anaconda
 +
 
 +
=== cuDNN ===
 +
The NVIDIA CUDA Deep Neural Network library (cuDNN) is a GPU-accelerated library of primitives for deep neural networks. cuDNN accelerates widely used deep learning frameworks, including Caffe2, MATLAB, Microsoft Cognitive Toolkit, TensorFlow, Theano, and PyTorch. If a specific version of cuDNN is needed, user can download from https://developer.nvidia.com/cudnn and choose '''"cuDNN [VERSION] Library for Linux (Power8/Power9)"'''.
 +
 
 +
The default cuDNN installed on the system is version 6 with CUDA-8 from IBM PowerAI. More recent cuDNN versions are installed as modules:
 +
<pre>
 +
cudnn/cuda9.0/7.0.5
 +
</pre>
 +
 
 +
=== Keras ===
 +
 
 +
Keras ([https://keras.io/ https://keras.io/]) is a popular high-level deep learning software development framework. It runs on top of other deep-learning frameworks such as TensorFlow.
 +
 
 +
*The easiest way to install Keras is to install Anaconda first, then install Keras by using using the pip command. Keras uses TensorFlow underneath to run neural network models. Before running code using Keras, be sure to load the PowerAI TensorFlow module and the cuda module.
 +
 
 +
*Keras can also be installed into a Python virtual environment by using '''pip'''. User can install optimized scipy (built with OpenBLAS) before installing Keras.
 +
In a virtual environment (python2.7 as example):
 +
<pre>
 +
pip install /scinet/sgc/Libraries/scipy/scipy-1.1.0-cp27-cp27mu-linux_ppc64le.whl
 +
pip install keras
 +
</pre>
 +
 
 +
=== NumPy/SciPy (built with OpenBLAS) ===
 +
 
 +
Optimized NumPy and SciPy are provided as Python wheels located in '''/scinet/sgc/Libraries/numpy''' and '''/scinet/sgc/Libraries/scipy''' and can be installed by '''pip'''. Please uninstall old numpy/scipy before installing the new ones.
 +
 
 +
=== PyTorch ===
 +
 
 +
PyTorch is the Python implementation of the Torch framework for deep learning.
 +
 
 +
It is suggested that you use PyTorch within Anaconda.
 +
 
 +
There is currently no build of PyTorch for POWER8-based systems. You will need to compile it from source.
 +
 
 +
Obtain the source code from here: [http://pytorch.org/ http://pytorch.org/]
 +
 
 +
Before building PyTorch, make sure to load cuda by running
 +
 
 +
<pre>
 +
module load cuda/8.0
 +
</pre>
 +
 
 +
NOTE: Do not have the gcc modules loaded when building PyTorch. Use the default version of gcc (currently v5.4.0) included with the operating system. Build will fail with later versions of gcc.
 +
 
 +
=== TensorFlow (new versions and python3) ===
 +
 
 +
The TensorFlow which is included in PowerAI may not be the most recent version. Newer versions of TensorFlow are provided as prebuilt Python Wheels that users can use '''pip''' to install under user space. Custom Python wheels are stored in '''/scinet/sgc/Applications/TensorFlow_wheels'''. It is highly recommended to install custom TensorFlow wheels into a Python virtual environment.
 +
 
 +
====Installing with Python2.7:====
 +
<div class="toccolours mw-collapsible mw-collapsed" style="overflow:auto;">
 +
* Create a virtual environment '''tensorflow-1.8-py2''' with packages installed with system:
 +
<pre>
 +
virtualenv --python=python2.7 --system-site-packages tensorflow-1.8-py2
 +
</pre>
 +
* Activate virtual environment:
 +
<pre>
 +
source tensorflow-1.8-py2/bin/activate
 +
</pre>
 +
* Install TensorFlow into the virtual environment: (A custom Numpy built with OpenBLAS library can be installed)
 +
<pre>
 +
pip install --upgrade --force-reinstall /scinet/sgc/Libraries/numpy/numpy-1.14.3-cp27-cp27mu-linux_ppc64le.whl
 +
pip install /scinet/sgc/Applications/TensorFlow_wheels/tensorflow-1.8.0-cp27-cp27mu-linux_ppc64le.whl
 +
</pre>
 +
</div>
 +
 
 +
====Installing with Python3.5:====
 +
<div class="toccolours mw-collapsible mw-collapsed" style="overflow:auto;">
 +
* Create a virtual environment '''tensorflow-1.8-py3''' with packages installed with system:
 +
<pre>
 +
virtualenv --python=python3.5 --system-site-packages tensorflow-1.8-py3
 +
</pre>
 +
* Activate virtual environment:
 +
<pre>
 +
source tensorflow-1.8-py3/bin/activate
 +
</pre>
 +
* Install TensorFlow into the virtual environment: (A custom Numpy built with OpenBLAS library can be installed)
 +
<pre>
 +
pip3 install --upgrade --force-reinstall /scinet/sgc/Libraries/numpy/numpy-1.14.3-cp35-cp35m-linux_ppc64le.whl
 +
pip3 install /scinet/sgc/Applications/TensorFlow_wheels/tensorflow-1.8.0-cp35-cp35m-linux_ppc64le.whl
 +
</pre>
 +
</div>
 +
 
 +
====Submitting jobs====
 +
<div class="toccolours mw-collapsible mw-collapsed" style="overflow:auto;">
 +
The above myjob.script file needs to be modified to run custom TensorFlow. '''cuda/9.0''' and '''cudnn/cuda9.0/7.0.5''' modules need to be loaded. Virtual environment needs to be activated.
 +
<pre>
 +
#!/bin/bash
 +
#SBATCH --nodes=1
 +
#SBATCH --ntasks=20  # MPI tasks (needed for srun)
 +
#SBATCH --time=00:10:00  # H:M:S
 +
#SBATCH --gres=gpu:4    # Ask for 4 GPUs per node
 +
 
 +
module purge
 +
module load cuda/9.0 cudnn/cuda9.0/7.0.5
 +
source tensorflow-1.8-py2/bin/activate #change this to the location where virtual environment is created
 +
 
 +
cd $SLURM_SUBMIT_DIR
 +
python code.py
 +
</pre>
 +
</div>
 +
 
 +
== LINKS ==
 +
 
 +
[https://www.olcf.ornl.gov/kb_articles/summitdev-quickstart/#System_Overview  Summit Dev System at ORNL]
 +
 
 +
== DOCUMENTATION ==
  
−
The PowerAI platform contains popular open machine learning frameworks such as Caffe, Tensorflow, and Torch. Run the <tt>module avail</tt> command for a complete listing. More information is available at this link: https://developer.ibm.com/linuxonpower/deep-learning-powerai/releases/. Release 4.0 is currently installed.
+
# GPU Cluster Introduction: [[Media:GPU_Training_01.pdf‎|SOSCIP GPU Platform]]
 +
-->

Latest revision as of 15:17, 5 October 2018


WARNING: SciNet is in the process of replacing this wiki with a new documentation site. For current information, please go to https://docs.scinet.utoronto.ca

SOSCIP GPU
S882lc.png
Installed September 2017
Operating System Ubuntu 16.04 le
Number of Nodes 14x Power 8 with 4x NVIDIA P100
Interconnect Infiniband EDR
Ram/Node 512 GB
Cores/Node 2 x 10core (20 physical, 160 SMT)
Login/Devel Node sgc01
Vendor Compilers xlc/xlf, nvcc

New Documentation Site

Please visit the new documentation site: https://docs.scinet.utoronto.ca/index.php/SOSCIP_GPU for updated information.

SOSCIP

The SOSCIP GPU Cluster is a Southern Ontario Smart Computing Innovation Platform (SOSCIP) resource located at theUniversity of Toronto's SciNet HPC facility. The SOSCIP multi-university/industry consortium is funded by the Ontario Government and the Federal Economic Development Agency for Southern Ontario [1].

Support Email

Please use <soscip-support@scinet.utoronto.ca> for SOSCIP GPU specific inquiries.


Retrieved from ‘https://oldwiki.scinet.utoronto.ca/index.php?title=SOSCIP_GPU&oldid=9425’