An Improved/Optimized Practical Non-Blocking PageRank Algorithm for Massive Graphs

• Published in: International journal of parallel programming Vol. 50; no. 3-4; pp. 381 - 404
• Main Authors: Eedi, Hemalatha, Karra, Sahith, Peri, Sathya, Ranabothu, Neha, Utkoor, Rahul
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.08.2022; Springer Nature B.V
• Subjects: Algorithms; Computer architecture; Computer memory; Computer Science; Feature extraction; Graphs; Iterative algorithms; Iterative methods; Microprocessors; Processor Architectures; Recommender systems; Search algorithms; Software Engineering/Programming and Operating Systems; Theory of Computation; Multi Threading; PageRank; Non-blocking mechanism; Barrier synchronization; Blocking mechanism; Shared memory architecture
• ISSN: 0885-7458; 1573-7640
• DOI: 10.1007/s10766-022-00725-6

PageRank kernel is a standard benchmark addressing various graph processing and analytical problems. The PageRank algorithm serves as a standard for many graph analytics and a foundation for extracting graph features and predicting user ratings in recommendation systems. The PageRank algorithm is an iterative algorithm that continuously updates the ranks of pages until it converges to a value. However, implementing the PageRank algorithm on a shared memory architecture while taking advantage of fine-grained parallelism with large-scale graphs is hard to implement. The experimental study and analysis of the parallel PageRank metric on large graphs and shared memory architectures using different programming models have been studied extensively. This paper presents the asynchronous execution of the PageRank algorithm to leverage the computations on massive graphs, especially on shared memory architectures. We evaluate the performance of our proposed non-blocking algorithms for PageRank computation on real-world and synthetic datasets using POSIX Multithreaded Library on a 56 core Intel(R) Xeon processor. We observed that our asynchronous implementations achieve 10 × to 30 × speed-up with respect to sequential runs and 5 × to 10 × improvements over synchronous variants.