Breadth first search vectorization on the Intel Xeon Phi

Abstract

Breadth First Search (BFS) is a building block for graph algorithms and has recently been used for large scale analysis of information in a variety of applications including social networks, graph databases and web searching. Due to its importance, a number of different parallel programming models and architectures have been exploited to optimize the BFS. However, due to the irregular memory access patterns and the unstructured nature of the large graphs, its efficient parallelization is a challenge. The Xeon Phi is a massively parallel architecture available as an off-the-shelf accelerator, which includes a powerful 512 bit vector unit with optimized scatter and gather functions. Given its potential benefits, work related to graph traversing on this architecture is an active area of research.We present a set of experiments in which we explore architectural features of the Xeon Phi and how best to exploit them in a top-down BFS algorithm but the techniques can be applied to the current state-of-the-art hybrid, top-down plus bottom-up, algorithms.We focus on the exploitation of the vector unit by developing an improved highly vectorized OpenMP parallel algorithm, using vector intrinsics, and understanding the use of data alignment and prefetching. In addition, we investigate the impact of hyperthreading and thread affinity on performance, a topic that appears under researched in the literature. As a result, we achieve what we believe is the fastest published top-down BFS algorithm on the version of Xeon Phi used in our experiments. The vectorized BFS topdown source code presented in this paper can be available on request as free-to-use software.

Vectorization of Hybrid Breadth First Search on the Intel Xeon Phi

“…Reference Approach optimisation vectorisation 2012 Saule and Ç atalyürek [17] top-down no optimisation automatic 2013 Gao et al [19] top-down bitmaps intrinsics 2013 Stanic et al [ The key contribution of this work builds on the studies carried out by Gao et al in [19] and [5]. In the first study, they present the vectorisation of the top-down BFS algorithm arXiv:1704.02259v2 [cs.DC] 20 Apr 2017 using vector intrinsic functions, which was outperformed by [15], which clarified the impact of prefetching, thread affinity and vector unit usage rate. The second study is related with the vectorisation of the hybrid BFS algorithm.…”

“…Particularly, the vectorisation of the hybrid involves the vectorised version of both algorithms. The vectorisation of the top-down algorithm is described and analysed in [15], whereas the vectorisation of the bottom-up is described further in Section 5. Table 2 shows an example of the switching points to swap between the top-down and the bottom-up approaches of the , , , ←getCounters() 14: swap(in, out) 15: ← 0 16: end while hybrid BFS algorithm for a graph created by the Graph 500 graph generator introduced in Section 6.…”

“…This paper investigates the optimisation and vectorisation of the BFS on the Xeon Phi, which is a parallel architecture containing advanced vector capabilities within the experimental framework of the Graph 500 benchmark. As a result, an novel optimised parallel version of the hybrid BFS is presented using vectorisation, building on top of the vectorised top-down BFS introduced by Paredes et al [15]. Note that CF'17, Siena, Italy 2017.…”

Exploring HPC and Big Data Convergence: A Graph Processing Study on Intel Knights Landing

“…(3) Low-level single node: These are single-node native platforms which are performance-oriented and include architecture-dependent optimization [47]. As a consequence, such implementations are able to exploit better the hardware capability, leveraging good performance, but the programming and tuning efforts are high and, more importantly, portability is difficult to achieve.…”

SlimSell: A Vectorizable Graph Representation for Breadth-First Search

“…There exist some BFS schemes for GPUs [31], [28], [11], [39], [14], [18]. However, they usually underutilize the available SIMD and vectorization parallelism as they focus on work-optimal traditional BFS or BFS based on SpMSpV that use fine-grained irregular memory accesses.…”