Performance analysis of graph processing frameworks

Abstract

Graphs have always been an interesting structure to study in both mathematics and computer science , and have become even more interesting in the context of online social networks, recommendation networks whose underlying network structures are nicely represented by graphs.The graphs are massive: Facebook social graph has billions of vertices and web graphs are much larger.With “large” graphs comes the desire to extract meaningful information from these graphs. In the age of multi-core CPUs and distributed computing, concurrent processing of graphs proves to be an important topic. Graph processing frameworks are being increasingly used to perform analysis on the enormous graphs like follower graphs in online social networks,web graph,recommendation graphs etc.Graphlab, FlashGraph, PowerGraph, X-stream are few frameworks are used to compute metrics such as pageank,shortest path etc on graphs. The lack of access locality when traversing edges makes it difficult to achieve good results in graph analysis. To gain an understanding of how graph processing frameworks perform, we conduct a study to experimentally compare Flash Graph and Graph lab Create using several metrics.The systems are compared with three different algorithms (Page Rank,weakly connected components,and Triangle counting) on single machine.Our evaluation shows that Graph lab create is performing better than Flash Graph.