Performance of Local Push Algorithms for Personalized PageRank on Multi-core Platforms

Personalized PageRank (PPR) is used to measure the importance of vertices with respect to a source vertex. PPR is a key kernel used in many real-world applications, such as information retrieval, recommendations, knowledge discovery, etc. Local push algorithms have been widely used for developing state-of-the-art fast PPR algorithms. In this paper, we analyze the computational characteristics of local push algorithms at the algorithm (error tolerance, damping factor) and hardware architecture (available memory, cache features) levels. First, we profile the algorithm to understand the effect of various algorithm parameters. We study the trade-offs between latency and accuracy of local push algorithms using various performance metrics including Top-K accuracy and scalability. Then, we perform our analysis on two state-of-the-art multi-core platforms to understand the latency of the algorithm for PPR computation on a single source vertex and its scalability for multiple vertices using thread-level parallelism. We analyze the impact of error tolerance and damping factor on the overall performance of the algorithms.