TuFast: A Lightweight Parallelization Library for Graph Analytics

Recently, there has been significant interest in large-scale graph analytics systems. However, most of the design efforts focus on accelerating graph analytics on giant graphs and/or in a distributed environment. Little attention focuses on the programmer usability perspective, which is critical to implementing ad-hoc analytics on moderate size graphs. In this paper, we present a lightweight transactional memory (TM) library TuFast which provides easy-to-use primitives for the end-user to agilely develop fast shared memory graph parallelization on a multi-core server. TuFast exploits recent CPU instructions set Hardware Transactional Memory (HTM), which has been available in off-the-shelf CPUs. HTM offers free transactional semantic but also suffers from capacity limitation. Our framework resolves the capacity challenge and efficiently utilizes HTM on graph parallelization by exploiting the graph degree information. Large scale graphs have a power-law degree distribution: a large proportion of the vertices with a small degree, fits in single HTM transactions; a small proportion of vertices with a big degree fits a pessimistic approach like locking; other vertices with a moderate degree can be processed with an optimistic approach with HTM acceleration. Our hybrid approach automatically adapts to the degree of graphs dynamically during the processing. The graph analytical jobs expressed via our library are straightforward and concise and outperform state-of-the-art distributed and multi-core graph analytical systems by up to 4 orders of magnitude.