Apache Spark for Scientific Data at Scale

Modern, robust, general-purpose cluster computing for data science

Presentation at SEA Class and Hands-on Workshop, 2017-06-22

Neal McBurnett Independent Consultant


  • Your name
  • Your association with UCAR
  • What brings you here
  • What you're hoping to get out of the next 2 days
  • A problem you might use Spark or TensorFlow for
  • Experience (Raise hands):
    • Code in Fortran? MPI? OpenMP? NCL? IDL? Python 2? Python 3? Other?
    • Tried Spark out? Used it for work?
    • Tried TensorFlow out? Used it for work?


  • Apache Spark is a productive tool for many scientific applications
    • Easier to write Spark code than MPI, OpenMP
    • Programming model is more expressive and helps compilers optimize for the cluster
  • Builds on convergence of data analysis technologies:
    • Databases: SQL, much research into query optimization
    • HPC: MPI, OpenMP, etc
    • Interactive exploratory data analysis: S, S++, R
    • Scala and Python knack for clear, expressive, high-level coding
    • MapReduce: Hadoop, Spark, with huge investment from business world
  • Related APIs and projects support many other applications
  • Many Scientific use cases

Note: I welcome questions, comments, corrections anytime

Apache Spark: popular, modern, elegant, productive

  • The largest open source project in data processing (in 2015. Still?)
    • 365 k Spark Meetup members worldwide
    • Over 750 contributors from 200+ organizations, still growing
    • Big turnout for our free online edX MOOC in 2015
    • IBM helping to "educate one million data scientists and data engineers on Spark"
  • APIs for Python (2 or 3), SQL, Java, Scala, R
  • Evolved from Hadoop MapReduce
  • Brought in-memory processing for fast iteration, then much more
  • Handles distributing data to nodes for you, and flexible caching
  • Can read only the data needed from smart upstream data sources: "predicate pushdown" of filters

Spark components and ecosystem

Core concept: RDD (Resilient Distributed Dataset)

  • RDD abstraction (Resilient Distributed Datasets), core aspect of Spark programming model
    • Matei Zaharia ACM Doctoral Dissertation Award
    • New abstraction that led to Spark
  • RDD is a collection of data, partitioned across the cluster
    • Each partition can be processed in parallel
    • RDD definition includes lineage - how each RDD depends on others
    • In general, a Directed Acyclic Graph (DAG)
    • Immutable
  • Two types of RDD operations
    • Transformations: modify the DAG: define new RDDs based on extisting ones
      • filter, map, groupByKey, join, sort
    • Actions apply, optimize, schedule and execute the DAG, return values
      • count, reduce, lookup, save
  • RDDs provide fault tolerance via recalculation, avoiding data replication
  • RDDs can be cached to make them more persistent in memory, or temporarily saved to local disk

Example of a lineage of RDDs. Also shows the job stages. Boxes with solid outlines are RDDs. Partitions are shaded rectangles, in black if they are already in memory. In this case, stage 1’s output RDD (B) is already in RAM, so the scheduler runs stage 2 and then 3.

If a node dies, the partitions it was responsible for can be recalculated, knowing the lineage and the immutable inputs

Explore Your Runtime in Spark UI

Real-time graph of the DAGs in your jobs, stage progress, storage, environment, nodes

Spark code compiled into efficient execution plan for the cluster

  • Something like compilers optimizing for pipelined CPUs

  • Spark mashes functional programming up with SQL optimizing technology

    • Catalyst compiler for parallel computations on your cluster
  • Code describes Directed Acyclic Graph

  • Remember Lazy evalution - not evaluated until you perform an action
  • Compiler can use full DAG, desired action, and even data contents, to make an optimized execution plan
  • Lack of side effects in functional programming enormously useful in parallel processing
  • Review and make use of your abstract algebra classes!
  • Lazy evaluation makes exploration, debugging quicker

Even higher-level API: Datasets and DataFrames

  • DataFrame API: Higher level than RDD, inspired by Pandas and R data frames, combined with SQL engine
  • More optimizations possible than with RDD
  • Unified with more general Datasets API in Spark 2.0: DataFrame = Dataset[Row]
  • Internally leverage the Spark SQL logical optimizer. Intelligently plan the physical execution of operations to work well on large datasets. This planning permeates all the way into physical storage, where optimizations such as predicate pushdown are applied based on analysis of user programs.
  • Create Datasets directly from Hive tables, Parquet files, Spark SQL, etc.

Even More Higher-level APIs

You can do lots of things with your data in Spark

  • ML - elegant API for machine learning, inspired by scikit-learn pipelines
  • MLlib - lower level Spark Machine Learning API, with some underlying implementations and more specialized algorithms
  • Spark SQL makes Spark accessible to many more people: analysts, managers
  • GraphX to bring graph processing to big data sets
  • Streaming, unified with batch processing

Scientific use cases for Spark

  • Researchers exploring data interactively with tools like Jupyter, IPython, Scala, Zeppelin, Databricks
    • The same code can be evolved into an efficient production system
    • See e.g. SciSpark from NASA: Automated Search for Mesoscale Convective Complexes
  • Machine learning at scale
  • Combining streaming and batch processing for real-time data products (weather forecasts?)
  • Pre-process datasets: Extract, Transform, Load (ETL) with modern tools
    • NASA use of MapReduce in Hadoop for MERRA/AS (Modern Era Retrospective-Analysis for Research and Applications Analytic Services). Augments netCDF files with indexes in HDFS. Could integrate with Spark.
  • Post-process data
    • Stream simulation results into a Spark analysis pipeline
    • Parallelize and process results from other tools: scikit-learn, Pandas, anything you can run on a node
    • Hyperparameter tuning for models using grid search with cross validation, e.g. with deep learning / Tensorflow via new Databricks code
  • Rethinking paper with case studies in astronomy (Spark-mAdd) and genomics (ADAM): faster than HPC state-of-the-art for some tasks, organizing around data formats like Parquet, designed for parallelization and compression.

Practicalities and Challenges

  • Most of the runtime and libraries are Scala or Java, cleverly integrated with Python
  • System is running on Java Virtual Machine behind the scenes
  • Error messages are often confusing
  • If partition in an RDD is larger than a node's memory, it spills to disk, can trash performance
    • So factor into application design, or have SSDs on your nodes, or size your cluster to match your data

"HPC systems are the mirrored image of data centers.

In a data center, the file system is optimized for latency (with local disks) and the network is optimized for bandwidth.

In a supercomputer, the file system is optimized for bandwidth (without local disk) while the network is optimized for latency." - IPCC

Not the Intergovernmental Panel on Climate Change....

The other IPCC: Intel Parallel Computing Center at Lawrence Berkeley Lab

IPCC is working to adapt Spark, which came from the world of data center clusters, to the supercomputer world.

Note: Yellowstone: no local disks on nodes

SDSC Comet (via XSEDE) does have local disks, for workloads that benefit

  • Apache Parquet: columnar storage format with great parallelization, compression, flexibility
  • Ceph: highly parallel object storage ala Amazon S3, along with other capabilities.
  • Apache Arrow: columnar in-memory data layer between Spark, Pandas, Parquet, Drill, etc: reduce serialization overhead
  • Apache Beam: Batch and strEAMing dataflow model and API from Google.
  • TensorFlow: open source library for numerical computation using data flow graphs
  • Apache Giraph: iterative graph processing, with similarities to MPI


  • Spark has many applications for scientific computing
  • It helps you be more productive at writing code and analyzing data
  • Spark optimizers make your code faster
  • Popular due to speed, flexibility, blend of interactive investigative analytics, batch data analytics, streaming updates
  • Lots of training opportunities and experience out there due to popularity
  • Runs in lots of environments, easier to develop on laptop, test on small cluster, redeploy on big one or in the cloud

Thanks to:

  • Insights and motivation from the Boulder Data Detectives LinkedIn Group (Wednesdays at BJ's)
  • Deep insights from Sameer Farooqui, Brian Clapper, and the rest of the team at Databricks Training
  • Great examples of Spark applications in climate science by Anderson Banihirwe
  • Zebula Sampedro, Dan Milroy and Davide Del Vento for original spark-setup-scripts, for running Spark on Janus and Yellowstone

Explore more.... Questions?     Neal McBurnett     http://bcn.boulder.co.us/~neal/