GraphLab: A framework for Parallel Machine Learning Documentation

1.0

Introduction

GraphLab is a powerful new system for designing and implementing parallel algorithms in machine learning. While the current targets multi-core shared memory parallel systems we are in the process of implementing a distributed version and plan to provide support for alternative parallel architectures include GPUs in the near future.
The easiest way to pick up GraphLab is to code! Here is a pagerank example which will provide you with some of the high level ideas of GraphLab And here is a more detailed example which provides more details as well as the the supporting APIs surrounding GraphLab.
The key pages of interest are:

• The graphlab::graph data structure.
  represents a directed graph container and is used extensively throughout GraphLab.
• Scopes
• Schedulers
• Shared Data

GraphLab is heavily templatized and the following two structures help to simplify template usage.

• The graphlab::core data structure.
  This provides a convenient wrapper around most of Graphlab.
• graphlab::types
  This provides typedefs for all shared memory GraphLab types.

GraphLab is additionally supported by a serialization library, a fast parallel/thread-safe random number generator, as well as a flexible command line parsing system.

• Serialization
• Random Number Generators
• graphlab::command_line_options
  The detailed example provides a good example of how this is used

The GraphLab library also has a collection of parallel utility classes and functions which may be useful.

Distributed GraphLab

There is a basic usage documentation for the distributed GraphLab implementation here: Using Distributed GraphLab .
It relies on a nice RPC implementation documented here: GraphLab RPC .