How to do Image Processing with GPUs


To demonstrate the capability of running a distributed job in PySpark using a GPU, this example uses NumbaPro and the CUDA platform for image analysis. This example executes 2-dimensional FFT convolution on images in grayscale and compares the execution time of CPU-based and GPU-based calculations.

Who is this for?

This how-to is for users of a Spark cluster who wish to run Python code using the YARN resource manager. This how-to will show you how to integrate third-party Python libraries with Spark.

Before you start

For this example, you’ll need Spark running with the YARN resource manager. You can install Spark and YARN using an enterprise Hadoop distribution such as Cloudera CDH or Hortonworks HDP.

You will also need valid Amazon Web Services (AWS) credentials in order to download the example data.

For this example, we recommend the use of the GPU-enabled AWS instance type g2.2xlarge and the AMI ami-12fd8178 (us-east-1 region), which has CUDA 7.0 and the NVIDIA drivers pre-installed. An example profile (to be placed in ~/.acluster/profiles.d/gpu_profile.yaml) is shown below:

name: gpu_profile
node_id: ami-12fd8178  # Ubuntu 14.04, Cuda 7.0, us-east-1 region
node_type: g2.2xlarge
num_nodes: 4
provider: aws_east
user: ubuntu

To execute this example, download the: example script or spark-numbapro.ipynb example notebook.

If you wish to use the spark-numbapro.ipynb example notebook the Jupyter Notebook plugin can be installed on the cluster using the following command:

acluster install notebook

Once the Jupyter Notebook plugin is installed, you can view Jupyter Notebook in your browser using the following command:

acluster open notebook

Install dependencies

If you have permission to install packages with acluster you can install the required packages on all nodes using the following command.

acluster conda install scipy matplotlib numbapro PIL

Load data into HDFS

First, we will load the sample text data into the HDFS data store. The following script will transfer sample image data (approximately 1.1 GB) from a public Amazon S3 bucket to the HDFS data store on the cluster.

Download the script to your cluster and insert your Amazon AWS credentials in the AWS_KEY and AWS_SECRET variables.

import subprocess


s3_path = 's3n://{0}:{1}@blaze-data/dogs-cats-img/images'.format(AWS_KEY, AWS_SECRET)
cmd = ['hadoop', 'distcp', s3_path, 'hdfs:///tmp/dogs-cats']

Note: The hadoop distcp command might cause HDFS to fail on smaller instance sizes due to memory limits.

Run the script on the cluster.


After a few minutes, the image data will be in the HDFS data store on the cluster and ready for analysis.

Running the Job

Run the script on the Spark cluster using spark-submit. The output shows the image processing execution times for the CPU-based vs. GPU-based calculations. Using Spark's default log4j profile: org/apache/spark/
15/11/09 02:33:21 INFO SparkContext: Running Spark version 1.5.1


15/11/09 02:33:45 INFO TaskSetManager: Finished task 6.0 in stage 1.0 (TID 13)
in 106 ms on ip-172-31-9-24.ec2.internal (7/7)
15/11/09 02:33:45 INFO YarnScheduler: Removed TaskSet 1.0, whose tasks have
all completed, from pool
15/11/09 02:33:45 INFO DAGScheduler: ResultStage 1
(collect at /tmp/anaconda-cluster/ finished in 4.844 s
15/11/09 02:33:45 INFO DAGScheduler: Job 1 finished:
collect at /tmp/anaconda-cluster/, took 4.854970 s

10 images
CPU: 6.91735601425
GPU: 4.88133311272


15/11/09 02:34:27 INFO TaskSetManager: Finished task 255.0 in stage 3.0 (TID 525)
 in 139 ms on ip-172-31-9-24.ec2.internal (256/256)
15/11/09 02:34:27 INFO YarnScheduler: Removed TaskSet 3.0, whose tasks have
all completed, from pool
15/11/09 02:34:27 INFO DAGScheduler: ResultStage 3
(collect at /tmp/anaconda-cluster/ finished in 19.340 s
15/11/09 02:34:27 INFO DAGScheduler: Job 3 finished:
collect at /tmp/anaconda-cluster/, took 19.400670 s

500 images
CPU: 22.1282501221
GPU: 19.8209779263


If something goes wrong consult the FAQ / Known issues page.

Further information

See the Spark and PySpark documentation pages for more information.

For more information on NumbaPro see the NumbaPro documentation.