GPOP: A scalable cache- and memory-efficient framework for Graph Processing Over Partitions Kartik Lakhotia Sourav Pati Rajgopal Kannan Viktor Prasanna University of Southern California {klakhoti, spati, rajgopak, prasanna}@usc.edu Abstract Past decade has seen the development of many shared-memory graph processing frameworks, intended to reduce the effort of developing high performance parallel applications. However many of these frameworks, based on Vertex-centric or Edge-centric paradigms suffer from several issues, such as poor cache utilization, irregular memory accesses, heavy use of synchronization primitives and theoretical inefficiency, that deteriorate overall performance and scalability. Recently, we proposed a cache and memory efficient partition-centric paradigm for computing PageRank [26]. In this paper, we generalize this approach to develop a novel Graph Processing Over Partitions (GPOP) framework that is cache-efficient, scalable and work-efficient. GPOP induces locality in memory accesses by increasing granularity of execution to vertex subsets called ’partitions’, thereby dramatically improving the cache performance of a variety of graph algorithms. It achieves high scalability by enabling completely lock and atomic free computation. GPOP’s built-in analytical performance model enables it to use a hybrid of source and partition- centric communication modes in a way that ensures work-efficiency each iteration, while simul- taneously boosting high bandwidth sequential memory accesses. Finally, the GPOP framework is designed with programmability in mind. It completely abstracts away underlying parallelism and programming model details from the user, and provides an easy to program set of APIs with the ability to selectively continue the active vertex set across iterations. Such functionality is useful for many graph algorithms but not intrinsically supported by the current frameworks. We extensively evaluate the performance of GPOP for a variety of graph algorithms, using several large datasets. We observe that GPOP incurs up to 9×, 6.8× and 5.5× less L2 cache misses compared to Ligra, GraphMat and Galois, respectively. In terms of execution time, GPOP is upto 19×, 9.3× and 3.6× faster than Ligra, GraphMat and Galois respectively. 1 Introduction Real world problems arising in web and social networks, transportation networks, biological systems etc. are often modeled as graph computation problems. Applications in these domains generate huge amounts of data that require efficient large-scale graph processing. To this purpose, many distributed frameworks have been proposed to process very large graphs on clusters [30, 22, 28, 11]. However, because of the high communication overheads of distributed systems, even single threaded implementations of graph algorithms have been shown to outperform many such frameworks running on several machines [34]. The growth in DDR capacity also allows large graphs to fit in the main memory of a single server. Consequently, many frameworks have been developed for high performance graph analytics 1 arXiv:1806.08092v3 [cs.DC] 19 Nov 2019