Skip to content

Latest commit

 

History

History
44 lines (35 loc) · 2.64 KB

README.md

File metadata and controls

44 lines (35 loc) · 2.64 KB
Raw

Stivo: An Embedded DSL for Distributed Data Parallel Processing

Stivo is a multy-stage programming approach for distributed data parallel processing. It is embedded in Scala, has a declarative and high level programming model similar to the Spark framework. However, due to the multi-stage programming approach, Stivo does the following optimizations:

  • Inserts projection before barriers to remove unused fields of Scala case classes.
  • Move pure expressions (e.g. Date, java.util.regex.Pattern) out of the hot path
  • Fuse all flatMap, map and filter and apply optimizations across them

From the same programming model, Stivo can generate code for Spark, Scoobi and to Crunch.

NOTE: Stivo is still in the experimental phase of development. Before making it production ready we plan to introduce more relational optimizations (moving filters, join reordering etc.) and make it interoperable with other DSLs like Regular Expressions DSL.

####Folder layout:

  • dsl contains the implementation of the DSL
  • virtualization-lms-core is the underlying framework, which deals with code motion and many other aspects of a compiler
  • scoobi is the target for generated sources for scoobi
  • spark is the target for generated sources for spark
  • crunch is the target for generated sources for crunch
  • pig contains the pig scripts we use as reference
  • benchmarking contains some scripts to help with benchmarking
  • utils contains small utilities
  • project contains the sbt configuration

This library is not (yet) meant for public usage. It is more of a prototype than a correct implementation.

####Getting started: The build should be reproducible now. For building you will need:

  • sbt 0.11.3 launcher
  • git

To get started, run setup.sh. This will fetch spark and deploy it to the local repository. Make sure that spark has been deployed correctly. When this is done, you can start sbt in the top distributed folder. For generating the classes you can use: ~ ; dsl/test ; dotgen ; gens/compile This will regenerate all test classes, update the utils, generate dot files and compile the generated files.

For benchmarking and packaging, I use the scripts in benchmarking. They are pretty messy though. Install Cloudera cdh3u4 on your machine to test the generated crunch and scoobi programs.

####Writing a program: See the examples in dsl/src/test/scala. To import your own class, create it in the dsl/structs folder, following the examples. Then use the dsl/startLift.sh script to refresh ApplicationOps, which you can then mix into your application.