Parallelization of a large scale density-friendly graph decomposition algorithm

Abstract

Nowadays on-line Social Network has lead to the creation of an enormous quantity of data, this data can be viewed as a graph composed by nodes that are people and edges that are the interactions between them. More scientists have started to study these structure, and one of the objective of these studies is the detection of the communities inside it, people whose interactions form a dense subgraph. There are several different strategies and algorithms for finding them, but in this thesis we focus in particular in one strategy based on a definition for subgraphs to be locally-dense and then that they form a nested chain decomposition of the graph sorted by increasing of density. In this algorithm we follow the implementation of a different solution based on the classic Frank-Wolfe algorithm that worked also for all the types of graphs of different sizes. We started from this basis and improving it with some refinements with the help of different analysis on the performance, time and on the cache we have created a parallel solution addressing both CPUs and GPUs. We show our experiments with a focus on the code, on the way that brought us to that solution and finally we show the result comparing all the solving techniques.