********************************************************************************
*     THIS IS A WORK IN PROGRESS.  DO NOT USE THIS FOR PRODUCTION WORK         *
********************************************************************************

VertexAPI2
==========


Introduction
------------

VertexAPI2 is a CUDA/C++ library for developing large graph analytics. The goal
is to hide the complexity of GPU programming while presenting a simple API.
VertexAPI2 uses the Gather-Apply-Scatter model used in GraphLab [cite].

VertexAPI2 is a work in progress. At this time we have a few working examples
like breadth first search, single source shortest paths and Page rank. The API
changes constantly as we evaluate and understand the best way to develop
further. If you are interested in using our project in your own work, please
send an email to the authors.


Known Issues
------------

The moderngpu library uses only one dimensional grids and consequently this
code will fail on Fermi cards for graphs with more than 8.3 million edges.


Installation
------------

The API is available as header files in the vertexAPI2 directory. To build the
example programs and a few utility objects, run 'make' in the vertexAPI2
directory.

Notes:
- A working installation of CUDA >= 5.0 is required.
- The moderngpu library should be available alongside the vertexAPI2 directory
  (mgpu patches to be made public on our repo?)


Building GraphLab regressions
-----------------------------

To perform a regression test against GraphLab, a working installation of
GraphLab is required. Edit the file
    vertexAPI2/PowerGraphReferenceImplementations/Makefile
to reflect the location of GraphLab on your system and then run 'make'
in the PowerGraphReferenceImplementations directory.  To download some
test graphs, run 'make' in the test-graphs directory.  Finally run 'make'
in the regressions directory.  At the end you can run './report.py' to
summarize the timing comparisons between GraphLab and vertexAPI2.


Credits:
--------

VertexAPI2 is developed by Royal Caliber LLC. as part of the DARPA XDATA
program under AFRL contract [enter contract number(s)].

Developers:
  Erich Elsen, Royal Caliber LLC, [email protected]
  Vishal Vaidyanathan, Royal Caliber LLC, [email protected]

Contributors:
  John Owens, UC Davis
  Yangzihao Wang, UC Davis

Acknowledgements:
  Sean Baxter, NVIDIA for the moderngpu library
  Duane Merrill, NVIDIA for the CUB library
  Joseph Gonzalez, Berkeley, for many informative discussions about the GAS
     abstraction and GraphLab.