Hprc banner tamu.png

Difference between revisions of "SW:TensorFlow"

From TAMU HPRC
Jump to: navigation, search
(Access)
(TensorFlow on Ada (CPU-only))
Line 23: Line 23:
 
  [NetID@ada ~]$ '''conda info --envs'''
 
  [NetID@ada ~]$ '''conda info --envs'''
  
===TensorFlow on Ada (CPU-only)===
+
==TensorFlow on Ada (CPU-only)==
 
A single version of TensorFlow is currently available on Ada. This version is limited to the CPU only (no GPU).  
 
A single version of TensorFlow is currently available on Ada. This version is limited to the CPU only (no GPU).  
  

Revision as of 13:17, 6 February 2018

TensorFlow

Description

TensorFlow is an open source software library for numerical computation using data flow graphs.

Access

TensorFlow is open to all HPRC users.

Please note that the GPU version of TensorFlow does not work on Ada. The GPU version must be used on Terra, which requires a different account and SUs than Ada. The CPU-only version does work on Ada.

There are restrictions as to which version (GPU/CPU) of TensorFlow works on each cluster. Please note these restrictions in the following sections and plan your jobs accordingly.

Anaconda and TensorFlow Packages

TAMU HPRC currently supports the user of TensorFlow though the Anaconda modules. There are a variety of Anaconda modules available on Ada and Terra.

While several versions of Anaconda have some TensorFlow environment installed, it is simplest to use exactly the versions in the following sections.

You can learn more about the module system on our SW:Modules page.

You can explore the available Anaconda environments on a per-module basis using the following:

[NetID@ada ~]$ module load Anaconda/[SomeVersion]
[NetID@ada ~]$ conda info --envs

TensorFlow on Ada (CPU-only)

A single version of TensorFlow is currently available on Ada. This version is limited to the CPU only (no GPU).

This version can be run on any of the 64GB or 256GB compute nodes, but caution should be used when testing on the login nodes.

To load this version (python 3.5):

[NetID@ada ~]$ module load Anaconda/3-4.0.0
[NetID@ada ~]$ source activate tensorflow
[NetID@ada ~]$ [run your Python program accessing TensorFlow]
[NetID@ada ~]$ source deactivate

TensorFlow on Terra (GPU-only)

There is one version of tensorflow (1.3.0) on Terra for module Anaconda/3-5.0.0.1 (recommended).

To load tensorflow-cua (python 3.6.2):

[NetID@terra ~]$ module load Anaconda/3-5.0.0.1
[NetID@terra ~]$ source activate tensorflow-gpu-1.3.0
[NetID@terra ~]$ [run your Python program accessing TensorFlow]
[NetID@terra ~]$ source deactivate

There are two versions of tensorflow available on Terra: 0.12.1 and 1.2.1 for module Anaconda/3-4.2.0. Both versions needs to load cuDNN/5.1-CUDA-8.0.44 module to access GPUs. Virtual environment tensorflow-cuda provides tensorflow 0.12.1, while tensorflow-gpu-1.2.1 provides tensorflow 1.2.1.

To load tensorflow-cua (python 3.5):

[NetID@terra ~]$ module load Anaconda/3-4.2.0
[NetID@terra ~]$ source activate tensorflow-cuda
[NetID@terra ~]$ module load cuDNN/5.1-CUDA-8.0.44
[NetID@terra ~]$ [run your Python program accessing TensorFlow]
[NetID@terra ~]$ source deactivate

To load tensorflow-cua (python 3.6.1):

[NetID@terra ~]$ module load Anaconda/3-4.2.0
[NetID@terra ~]$ source activate tensorflow-gpu-1.2.1
[NetID@terra ~]$ module load cuDNN/5.1-CUDA-8.0.44
[NetID@terra ~]$ [run your Python program accessing TensorFlow]
[NetID@terra ~]$ source deactivate

Example TensorFlow Script

As with any job on the system, TensorFlow should be used via the submission of a job file. Scripts using TensorFlow are written in Python, and thus TensorFlow scripts should not be written directly inside a job file or entered in the shell line by line. Instead, a separate file for the Python/TensorFlow script should be created, which can then be executed by the job file.

To create a new script file, simply open up the text editor of your choice.

Below is an example script (entered in the text editor of your choice):

import tensorflow as tf
x = tf.constant(35, name='x')
y = tf.Variable(x + 5, name ='y')
model = tf.global_variables_initializer()
with tf.Session() as session:
       session.run(model)
       print(session.run(y))

It is recommended to save this script with a .py file extension, but not necessary.

Once saved, the script can be tested on a login node by entering:

[NetID@terra ~]$ python testscript.py

NOTE: Make sure to run this command from the same directory that the script is saved in.

NOTE: While acceptable to test programs on the login node, please do not run extended or intense computation on these shared resources. Use a batch job and the compute nodes for heavy processing.

NOTE: Multi-core TensorFlow scripts potentially use ALL available cores by default. This can inadvertently crash the login node. Multi-core TensorFlow scripts must be tested within a batch job.

Installing Additional Python/TensorFlow Packages

While multiple versions of Python, Anaconda, and TensorFlow are available on our clusters, it is at times desired to have some specialized libraries or packages installed in addition to our pre-installed software.

Software installations by HPRC staff is usually reserved for generalized/popular or complex installations. There is a 10-15 business day (2-3 week) turn-around time for software installation requests.

You are encouraged to save time by trying to install your own Python/TensorFlow packages via the process in the follow sections.

General Installation Notes

User installation of Python packages is straightforward except for a few conditions. The following installation notes comprise most issues users encounter.

  • Disk/File Quota: Storage within $HOME is limited. Packages should be installed in user's $SCRATCH. See Ada File Systems or Terra File Systems for more info.
  • Internet Connection: Only login nodes have access to the Internet. Attempting to use pip or other Internet access from a batch job (on a compute node) will fail.
  • Mixing Tools: myPython, Anaconda, and myAnaconda must be used exclusively of one another. Attempting to mix these in order to add a package/library will fail.
  • Version Compatibility: Some packages require specific/older version of TensorFlow to be installed. Please verify package compatibility against available versions.

Adding TensorFlow Packages: Terra

For TensorFlow packages that require installation:

[NetID@terra ~]$ ml purge
[NetID@terra ~]$ ml Python/3.5.2-intel-2017A
[NetID@terra ~]$ virtualenv tdenv
[NetID@terra ~]$ source tdenv/bin/activate
[NetID@terra ~]$ pip install [link to package file]

If an older version of TensorFlow is needed, the package should be installed with the following commands:

[NetID@terra ~]$ ml purge
[NetID@terra ~]$ ml Python/3.5.2-intel-2017A
[NetID@terra ~]$ virtualenv tdenv
[NetID@terra ~]$ source tdenv/bin/activate
[NetID@terra ~]$ pip install tensorflow==1.0
[NetID@terra ~]$ pip install [link to package file]

Usage on the Login Nodes

Please limit interactive processing to short, non-intensive usage. Use non-interactive batch jobs for resource-intensive and/or multiple-core processing. Users are requested to be responsible and courteous to other users when using software on the login nodes.

The most important processing limits here are:

  • ONE HOUR of PROCESSING TIME per login session.
  • EIGHT CORES per login session on the same node or (cumulatively) across all login nodes.

Anyone found violating the processing limits will have their processes killed without warning. Repeated violation of these limits will result in account suspension.
Note: Your login session will disconnect after one hour of inactivity.

Usage on the Compute Nodes

Non-interactive batch jobs on the compute nodes allows for resource-demanding processing. Non-interactive jobs have higher limits on the number of cores, amount of memory, and runtime length.

For instructions on how to create and submit a batch job, please see the appropriate wiki page for each respective cluster:

Ada Example Job Script

A single-core, single node example that requests a total of 5000MB of RAM: (Last updated January 26, 2017)

#BSUB -J tfJob               # sets the job name to tfJob.
#BSUB -L /bin/bash           # uses the bash login shell to initialize the job's execution environment.
#BSUB -W 2:00                # sets to 2 hours the job's runtime wall-clock limit.
#BSUB -n 1                   # assigns 1 core for execution.
#BSUB -R "span[ptile=1]"     # assigns 1 core per node.
#BSUB -R "rusage[mem=5000]"  # reserves 5000MB per process/CPU for the job (5GB * 1 Core = 5GB per node) 
#BSUB -M 5000		     # sets to 5,000MB (~5GB) the per process enforceable memory limit.
#BSUB -o TensorFlow.%J       # directs the job's standard output to TensorFlow.jobid


# Load the modules
module load Anaconda/3-4.0.0 

# Activate TensorFlow library
source activate tensorflow

# Launch your Python program
[run your Python program accessing TensorFlow]

# Deactivate TensorFlow ibrary
source deactivate

To submit the batch job, run:

[NetID@ada ~]$ bsub < jobscript

Terra Example

Usage on the VNC Nodes

The VNC nodes allow for usage of the a graphical user interface (GUI) without disrupting other users.

VNC jobs and GUI usage do come with restrictions. All VNC jobs are limited to a single node (Terra: 28 cores/64GB). There are fewer VNC nodes than comparable compute nodes.

For more information, including instructions, on using software on the VNC nodes, please visit our Terra Remote Visualization page.

Short Course

HPRC hosts a TensorFlow short course once per semester. Information can be found HERE.

The slide deck associated with this TensorFlow short course can be found HERE.