Section: New Software and Platforms
Participants : Aurélien Lancin, Paul Bertot, Nicolas Chleq [SED-SOP] , David Coudert, Luc Hogie, Fabrice Huet [Scale] , Flavian Jacquot, Arnaud Legout [Diana] , Eric Madelaine [Scale] , Michel Syska [coordinator] .
The objective of BigGraphs is to provide a distributed middleware for very large graphs processing. This new project has received a development grant (ADT) from Inria and is a joint work of three EPI from Inria: COATI, DIANA and SCALE.
The first phase of the project consists in the evaluation of the existing middlewares such as GraphX/Spark or Giraph/Hadoop with respect to the following criteria: ease of deployment, maintenance and use; variety of programming models (Map/Reduce, BSP, (a)synchronous message passing, centralized programming, mobile agent-based, etc.); overall efficiency and memory footprint; etc. One of the chosen use cases is a subgraph of the Twitter graph with 3 millions of nodes and 200 millions of edges. The experiments are run on the NEF cluster at Inria. We have implemented and tested the classic algorithms (using the BSP model): page rank, BFS, connected components as well as the iFUB algorithm for computing the diameter of large graphs.
In parallel, we are testing new ideas through the development of custom solutions for the deployment of application code in heterogeneous environments, for automatic discovery of cluster architecture, for the design of distributed object oriented applications, and techniques for distributed graph computing (asynchronous BSP, messaging, multi-core parallelism, etc.).
The next phase is to decide whether one framework is matching our needs (and use it as a basis for further developments) or if we have to produce our own.