Overview

In this tutorial, we see how to submit a tensorflow job on the OSG through Singularity containers. We currently offer CPU and GPU containers for tensorflow (both based on Ubuntu). Here, we focus on CPU container.

Tutorial files

Let us utilize the tutorial command. In the command prompt, type

 $ tutorial tensorflow-matmul  (Copies input and script files to the directory tutorial-tf-matmul)

This will create a directory tutorial-tensorflow-matmul with the following files

tf_matmul.py            (Python program to multiply two matrices using tensorflow package)
tf_matmul.submit        (HTCondor Job description file)
tf_matmul_wrapper.sh    (Job wrapper shell script that executes the python program)

Matrix multiplication with tensorflow

The Python program tf_matmul.py uses tensorflow to perform the matrix multiplication of a 2x2 matrix. Indeed, this is not the best use case of tensorflow. This example is just fine to see how to submit the tensorflow job on the OSG.

Executing the script inside the singularity container

Before running this job on the OSG, let us see how to execute the tensorflow example on the submit host. Execute the Python program in the shell prompt

$ python tf_matmul.py

Traceback (most recent call last):
  File "tf_matmul.py", line 3, in <module>
    import tensorflow as tf
ImportError: No module named tensorflow

The error message says that tensorflow is not available.

We need to execute the program inside the tensorflow container. Singularity offers couple of ways to run an image. One of them is to execute a shell inside the image (See Singularity documentation for more details).

$ singularity shell /cvmfs/singularity.opensciencegrid.org/opensciencegrid/tensorflow:latest

This should drop you inside the container shell in few minutes. The tensorflow image tensorflow:latest is located at /cvmfs/singularity.opensciencegrid.org/opensciencegrid/ (more details about image file construction and distribution are outlined here)

Now we run the program inside the container

$ python tf_matmul.py
result of matrix multiplication
===============================
[[  1.00000000e+00   0.00000000e+00]
 [ -4.76837158e-07   1.00000024e+00]]
===============================

This is a 2x2 matrix multiplication and should be done in a minute or two.

Now let us see how to run this Python program on the remote machine as a singularity containter job.

Note: You may see the warning

 2017-07-16 12:31:44.841458: W tensorflow/core/platform/cpu_feature_guard.cc:45] The TensorFlow library wasn't compiled to use SSE4.1 instructions, but these are available on your machine and could speed up CPU computations.

that is related to the optimization of tensorflow installation on specific architecture.

Job execution and submission files

We want to run the program on a remote worker machine on the OSG that supports the singularity container. So we set the requirement in our HTCondor description

Requirements = HAS_SINGULARITY == True

In addition, we also provide the full path of the image via the keyword +SingularityImage:

+SingularityImage = "/cvmfs/singularity.opensciencegrid.org/opensciencegrid/tensorflow:latest"

The image is distributed to the remote worker machines through cvmfs. Although there are multiple ways to aquire the image file for a job on the OSG machine, the image distributed through cvmfs is preferred.

Let us take a look at the condor job description file tf_matmul.submit:

# The UNIVERSE defines an execution environment. You will almost always use VANILLA.
Universe = vanilla

# These are good base requirements for your jobs on OSG. It is specific on OS and
# OS version, cores, and memory.
Requirements = HAS_SINGULARITY == True
request_cpus = 1
request_memory = 2 GB
request_disk = 4 GB

# Singularity settings
+SingularityImage = "/cvmfs/singularity.opensciencegrid.org/opensciencegrid/tensorflow:latest"

# EXECUTABLE is the program that your job will run. It's often useful
# to create a shell script to "wrap" your actual work.
Executable = tf-matmul-wrapper.sh
Arguments =

# inputs/outputs
transfer_input_files = tf_matmul.py
transfer_output_files =

# ERROR and OUTPUT are the error and output channels from your job
# that HTCondor returns from the remote host.
Error = $(Cluster).$(Process).error
Output = $(Cluster).$(Process).output

# The LOG file is where HTCondor places information about your
# job's status, success, and resource consumption.
Log = $(Cluster).$(Process).log

# Send the job to Held state on failure. 
on_exit_hold = (ExitBySignal == True) || (ExitCode != 0)

# Periodically retry the jobs every 1 hour, up to a maximum of 5 retries.
periodic_release =  (NumJobStarts < 5) && ((CurrentTime - EnteredCurrentStatus) > 60*60)

# QUEUE is the "start button" - it launches any jobs that have been
# specified thus far.
Queue 1

The wrapper script tf-matmul-wrapper.sh is pretty normal one which executes the python program.

#!/bin/bash
python tf_matmul.py > tf_matmul.output

Job submmision

We submit the job using condor_submit command as follows

$ condor_submit tf_matmul.submit

The job will look for a machine on OSG that has singularity installed, creates the singularity container with the image /cvmfs/singularity.opensciencegrid.org/opensciencegrid/tensorflow:latest and executes the program tf_matmul.py.

The present job should be finished quickly (less than an hour). You can check the status of the submitted job by using the condor_q command as follows

$ condor_q username  # The status of the job is printed on the screen. Here, username is your login name.

The output of the job is available in the file tf_matmul.output.

Running on GPUs

You can also steer the job to run on GPUs, but note that the number of GPUs available on OSG is limited. Even though the job will execute faster, it might sit in the queue waiting longer than a CPU-only job.

The submit file for a GPU jobs is tf_matmul_gpu.submit The only difference is request_gpus = 1 and specifying a GPU image:

request_gpus = 1
...
+SingularityImage = "/cvmfs/singularity.opensciencegrid.org/opensciencegrid/tensorflow-gpu:latest"

Getting Help

For assistance or questions, please email the OSG User Support team at user-support@opensciencegrid.org or visit the help desk and community forums.

 

This page was updated on Jun 22, 2018 at 08:01 from tutorials/tutorial-tensorflow-matmul/README.md.