This document summarizes a presentation about deep learning workflows and best practices on Apache Spark. It discusses how deep learning fits within broader data pipelines for tasks like training and transformation. It also outlines recurring patterns for integrating Spark and deep learning frameworks, including using Spark for data parallelism and embedding deep learning transforms. The presentation provides tips for developers on topics like using GPUs with PySpark and monitoring deep learning jobs. It concludes by discussing challenges in the areas of distributed deep learning and Spark integration.

1. Deep Learning on Apache®
Spark™: Workflows and Best
Practices
Tim Hunter (Software Engineer)
Jules S. Damji (Spark Community Evangelist)
May 4, 2017

2. Agenda
• Logistics
• Databricks Overview
• Deep Learning on Apache® Spark™: Workflows and Best
Practices
• Q & A

3. Logistics
• We can’t hear you…
• Recording will be available...
• Slides and Notebooks will be available...
• Queue up Questions ….
• Orange Button for Tech Support difficulties...

4. Empower anyone to innovate faster with big data.
Founded by the creators of Apache Spark.
Contributes 75%of the open source code,
10x more than any other company.
VISION
WHO WE ARE
A data processing for data scientists, data engineers, and data
analysts that simplifies that data integration, real-time
experimentation, machine learning and deployment of
production pipelines .
PRODUCT

5. A New Paradigm
SECOND GENERATION
THE BEST OF BOTH WORLDS
Hadoop + data lake
Hard to centralize data and
extract value with disparate tools
Virtual analytics
• Holisticallyanalyze data from
data warehouses, data lakes,
and other data stores
• Utilize a single engine for batch,
ML, streaming & real-time
queries
• Enable enterprise-wide
collaboration
+
FIRST GENERATION
Data warehouses
ETL process is rigid, scaling out
is expensive, limited to SQL

6. CLUSTER TUNING &
MANAGEMENT
INTERACTIVE
WORKSPACE
PRODUCTION
PIPELINE
AUTOMATION
OPTIMIZED DATA
ACCESS
DATABRICKS ENTERPRISE SECURITY
YOUR TEAMS
Data Science
Data Engineering
Many others…
BI Analysts
YOUR DATA
Cloud Storage
Data Warehouses
Data Lake
VIRTUAL ANALYTICS PLATFORM

7. Deep Learning on Apache®
Spark™: Workflows and Best
Practices
Tim Hunter (Software Engineer)
May 4 , 2017

8. About Me
• Tim Hunter
• Software engineer @ Databricks
• Ph.D. from UC Berkeley in Machine Learning
• Very early Spark user
• Contributor to MLlib
• Author of TensorFrames and GraphFrames

9. Deep Learning and Apache Spark
Deep Learning frameworks w/ Spark bindings
• Caffe (CaffeOnSpark)
• Keras (Elephas)
• MXNet
• Paddle
• TensorFlow(TensorFlowOnSpark,TensorFrames)
Extensions to Spark for specialized hardware
• Blaze (UCLA & Falcon Computing Solutions)
• IBM Conductor with Spark
Native Spark
• BigDL
• DeepDist
• DeepLearning4J
• MLlib
• SparkCL
• SparkNet

10. Deep Learning and Apache Spark
2016: the year of emerging solutions for Spark + Deep Learning
No consensus
• Many approaches for libraries: integrate existing ones with Spark, build on
top of Spark, modify Spark itself
• Official Spark MLlib support is limited(perceptron-like networks)

11. One Framework to Rule Them All?
Should we look for The One Deep Learning Framework?

12. Databricks’ perspective
• Databricks: hosted Spark platform on public cloud
• GPUs for compute-intensive workloads
• Customers use many Deep Learning frameworks: TensorFlow, MXNet, BigDL,
Theano, Caffe, and more
This talk
• Lessons learned from supporting many Deep Learning frameworks
• Multiple ways to integrate Deep Learning & Spark
• Best practices for these integrations

13. Outline
• Deep Learning in data pipelines
• Recurring patterns in Spark + Deep Learning integrations
• Developer tips
• Monitoring

14. Outline
• Deep Learning in data pipelines
• Recurring patterns in Spark + Deep Learning integrations
• Developer tips
• Monitoring

15. ML is a small part of data pipelines.
Hidden technical debt in Machine Learning systems
Sculley et al., NIPS 2016

16. DL in a data pipeline: Training
Data
collection
ETL Featurization Deep
Learning
Validation Export,
Serving
compute intensive IO intensiveIO intensive
Large cluster
High memory/CPU ratio
Small cluster
Low memory/CPU ratio

17. DL in a data pipeline: Transformation
Specialized data transforms: feature extraction & prediction
Input Output
cat
dog
dog
Saulius Garalevicius - CC BY-SA3.0

18. Outline
• Deep Learning in data pipelines
• Recurringpatterns in Spark + Deep Learning integrations
• Developer tips
• Monitoring

19. Recurring patterns
Spark as a scheduler
• Data-parallel tasks
• Data stored outside Spark
Embedded Deep Learning transforms
• Data-parallel tasks
• Data stored in DataFrames/RDDs
Cooperative frameworks
• Multiple passes over data
• Heavy and/or specialized communication

20. Streaming data through DL
Primary storage choices:
• Cold layer (HDFS/S3/etc.)
• Local storage: files, Spark’s on-disk persistence layer
• In memory: SparkRDDs or SparkDataFrames
Find out if you are I/O constrained or processor-constrained
• How big is your dataset? MNIST or ImageNet?
If using PySpark:
• All frameworks heavily optimized for diskI/O
• Use Spark’s broadcastfor small datasets that fitin memory
• Reading files is fast: use local files when it does not fit

21. Cooperative frameworks
• Use Spark for data input
• Examples:
• IBM GPU efforts
• Skymind’s DeepLearning4J
• DistML and other Parameter Server efforts
RDD
Partition 1
Partition n
RDD
Partition 1
Partition m
Black box

22. Cooperative frameworks
• Bypass Spark for asynchronous / specific communication
patterns across machines
• Lose benefit of RDDs and DataFrames and
reproducibility/determinism
• But these guarantees are not requested anyway when doing
deep learning (stochastic gradient)
• “reproducibility is worth a factor of 2” (Leon Bottou, quoted by
John Langford)

23. Outline
• Deep Learning in data pipelines
• Recurring patterns in Spark + Deep Learning integrations
• Developer tips
• Monitoring

24. The GPU software stack
• Deep Learning commonly used with GPUs
• A lot of workon Spark dependencies:
• Few dependencies on local machine when compiling Spark
• The build process works well in a largenumber of configurations (just scala +
maven)
• GPUs present challenges: CUDA, support libraries, drivers, etc.
• Deep softwarestack, requires careful construction (hardware+ drivers + CUDA
+ libraries)
• All these are expected by the user
• Turnkey stacks just starting to appear

25. • Provide a Docker image with all the GPU SDK
• Pre-install GPU drivers on the instance
Container:
nvidia-docker,
lxc, etc.
The GPU software stack
GPU hardware
Linux kernel NV Kernel driver
CuBLAS CuDNN
Deep learning libraries
(Tensorflow, etc.) JCUDA
Python / JVM clients
CUDA
NV kernel driver (userspace interface)

26. Using GPUs through PySpark
• Popular choice for many independent tasks
• Many DL packages have Python interfaces: TensorFlow,
Theano, Caffe, MXNet, etc.
• Lifetime for python packages: the process
• Requires some configuration tweaks in Spark

27. PySpark recommendation
• spark.executor.cores = 1
• Gives the DL framework full access over all the resources
• Important for frameworks that optimize processor pipelines

28. Outline
• Deep Learning in data pipelines
• Recurring patterns in Spark + Deep Learning integrations
• Developer tips
• Monitoring

29. Monitoring
?

30. Monitoring
• How do you monitor the progress of your tasks?
• It depends on the granularity
• Around tasks
• Inside (long-running) tasks

31. Monitoring: Accumulators
• Good to check throughput
or failure rate
• Works for Scala
• Limited use for Python
(for now, SPARK-2868)
• No “real-time” update
batchesAcc = sc.accumulator(1)
def processBatch(i):
global acc
acc += 1
# Process image batch here
images = sc.parallelize(…)
images.map(processBatch).collect()

32. Monitoring: external system
• Plugs into an external system
• Existing solutions: Grafana, Graphite, Prometheus, etc.
• Most flexible, but more complex to deploy

33. Conclusion
• Distributed deep learning: exciting and fast-moving space
• Most insights are specific to a task, a dataset and an algorithm:
nothing replaces experiments
• Get started with data-parallel jobs
• Move to cooperative frameworks only when your data are too large.

34. Challenges to address
For Spark developers
• Monitoringlong-running tasks
• Presentingand introspecting intermediate results
For DL developers
• What boundary to put between the algorithm and Spark?
• How to integrate with Spark at the low-level?

35. Resources
Recent blog posts — http://databricks.com/blog
• TensorFrames
• GPU acceleration
• Getting started with Deep Learning
• Intel’s BigDL
Docs for Deep Learning on Databricks — http://docs.databricks.com
• Getting started
• Spark integration

36. SPARK SUMMIT 2017
DATA SCIENCE AND ENGINEERING AT SCALE
JUNE 5 – 7 | MOSCONE CENTER | SAN FRANCISCO
ORGANIZED BY spark-summit.org/2017

37. Thank You!
Questions?
Happy Sparking & Deep Learning!