Skip to main content

Using Rapids cuDF in a Colab notebook

During last Spark+AI Summit Europe 2019 I had a chance to attend a talk from Miguel Martinez who was presenting Rapids, the new Open Source framework from NVIDIA for GPU accelerated end-to-end Data Science and Analytics.


Fig. 1 - Overview of the Rapids eco-system

Rapids is a suite of Open Source libraries:

  • cuDF
  • cuML
  • cuGraph
  • cuXFilter

I enjoied the presentation and liked the idea of this initiative, so I wanted to start playing with the Rapids libraries in Python on Colab, starting from cuDF, but the first attempt came with an issue that I eventually solved. So in this post I am going to share how I fixed it, with the hope it would be useful to someone else running into the same blocker.
I am assuming here you are already familiar with Google Colab. I am using Python 3.x as Python 2 isn't supported by Rapids.
Once you have created a new notebook in Colab, you need to check if the runtime for it is set to use Python 3 and uses a GPU as hardware accelerator. You can check and eventually set it from the toolbar menu Runtime -> Change runtime type. Then you have to check that Colab has given you access to a NVIDIA Tesla GPU, as it is supported by Rapids. You can do this check from the notebook code by running

!nvidia-smi

and look for the GPU type, as highligthed in figure 2:

Fig. 2 - The output of the nvidia-smi command in a Colab notebook

The next step would be the installation of the cuDF library for CUDA 10.x. You can use pip from the notebook:

!pip install cudf-cuda100

At the end of the installation, before importing the cuDF package, you need to do an extra setup that I am going to explain below. If you don't do it, any time you try to execute the cuDF API, you will receive the following misleading error:

Fig. 3 - The error message in Colab when trying to use the cuDF API

The message states that the cudatoolkit isn't installed in your environment, so the numba package cannot be find, which isn't true, as cudatoolkit is already part of the Colab environment. The real root cause of this error is that the paths to the drivers are different in the runtime from what expected by the library. So you have to locate them first

dev_lib_path = !find / -iname 'libdevice'
nvvm_lib_path = !find / -iname 'libnvvm.so'


and setup the environment accordingly

import os

if len(dev_lib_path) > 0:
    os.environ['NUMBAPRO_LIBDEVICE'] = dev_lib_path[0]
else:
    print('The device lib is missing.')
if len(nvvm_lib_path) > 0:
    os.environ['NUMBAPRO_NVVM'] = nvvm_lib_path[0]
else:
    print('NVVM is missing.')


I have also put extra checks in case the paths shouldn't be found, which would probably be useless, but you never know :)

Now you can import cuDF, numpy and any other Python package you would need

import cudf
import numpy as np
...


and start playing with GPU Data Frames

df = cudf.DataFrame()
...


Enjoy it!

I will share my impressions on this library and the others in Rapids as soon as I have completed some PoCs.

Comments

  1. Feedback from Miguel Martinez (NVIDIA): by installing cuDF using pip, you will get the release 0.61. In order to install a latter or the latest release you have to use conda onstead. Thanks Miguel for the heads up.

    ReplyDelete
  2. This video helps me to understand Matplotlib whats your opinion guys.

    ReplyDelete
  3. Nice blog, very informative content.Thanks for sharing, waiting for the next update…
    Why learn Python?
    Why is Python so popular?

    ReplyDelete
  4. Pretty Post! Thank you so much for sharing this good content, it was so nice to read and useful to improve my knowledge as an updated one, keep blogging.

    Python Certification Training in Electronic City

    ReplyDelete

Post a Comment

Popular posts from this blog

Streamsets Data Collector log shipping and analysis using ElasticSearch, Kibana and... the Streamsets Data Collector

One common use case scenario for the Streamsets Data Collector (SDC) is the log shipping to some system, like ElasticSearch, for real-time analysis. To build a pipeline for this particular purpose in SDC is really simple and fast and doesn't require coding at all. For this quick tutorial I will use the SDC logs as example. The log data will be shipped to Elasticsearch and then visualized through a Kibana dashboard. Basic knowledge of SDC, Elasticsearch and Kibana is required for a better understanding of this post. These are the releases I am referring to for each system involved in this tutorial: JDK 8 Streamsets Data Collector 1.4.0 ElasticSearch 2.3.3 Kibana 4.5.1 Elasticsearch and Kibana installation You should have your Elasticsearch cluster installed and configured and a Kibana instance pointing to that cluster in order to go on with this tutorial. Please refer to the official documentation for these two products in order to complete their installation (if you do

Exporting InfluxDB data to a CVS file

Sometimes you would need to export a sample of the data from an InfluxDB table to a CSV file (for example to allow a data scientist to do some offline analysis using a tool like Jupyter, Zeppelin or Spark Notebook). It is possible to perform this operation through the influx command line client. This is the general syntax: sudo /usr/bin/influx -database '<database_name>' -host '<hostname>' -username '<username>'  -password '<password>' -execute 'select_statement' -format '<format>' > <file_path>/<file_name>.csv where the format could be csv , json or column . Example: sudo /usr/bin/influx -database 'telegraf' -host 'localhost' -username 'admin'  -password '123456789' -execute 'select * from mem' -format 'csv' > /home/googlielmo/influxdb-export/mem-export.csv