Traversing large graphs on GPUs with unified memory

Abstract

Due to the limited capacity of GPU memory, the majority of prior work on graph applications on GPUs has been restricted to graphs of modest sizes that fit in memory. Recent hardware and software advances make it possible to address much larger host memory transparently as a part of a feature known as unified virtual memory. While accessing host memory over an interconnect is understandably slower, the problem space has not been sufficiently explored in the context of a challenging workload with low computational intensity and an irregular data access pattern such as graph traversal. We analyse the performance of breadth first search (BFS) for several large graphs in the context of unified memory and identify the key factors that contribute to slowdowns. Next, we propose a lightweight offline graph reordering algorithm, HALO (Harmonic Locality Ordering), that can be used as a pre-processing step for static graphs. HALO yields speedups of 1.5x-1.9x over baseline in subsequent traversals. Our method specifically aims to cover large directed real world graphs in addition to undirected graphs whereas prior methods only account for the latter. Additionally, we demonstrate ties between the locality ordering problem and graph compression and show that prior methods from graph compression such as recursive graph bisection can be suitably adapted to this problem.

One Algorithm to Evaluate Them All: Unified Linear Algebra Based Approach to Evaluate Both Regular and Context-Free Path Queries

“…Evaluation of Azimov's algorithm shows that it is possible to improve performance by using GPGPU because operations of linear algebra can be efficiently implemented on GPGPU (Mishin et al, 2019;Terekhov et al, 2020). Moreover, for practical reasons, it is interesting to provide a multi-GPU version of the algorithm and to utilize unified memory, which is suitable for linear algebra based processing of out-of-GPGPU-memory data and traversing on large graphs (Chien et al, 2019;Gera et al, 2020).…”

Cache-Efficient Fork-Processing Patterns on Large Graphs

“…We evaluate competing systems' performance on three applications, BC, LL, and NCP. To keep consistent with previous work [1,23,53], we configure the three applications as follows.…”

“…The algorithm samples the starting nodes from the graph. Therefore, in our evaluation, we randomly sample a batch of 100 source vertices for each graph [23].…”

“…Next, the algorithm gathers the results of each BFS to obtain the centrality of vertices [10]. Although various algorithm variants have been proposed, they have common FPPs of launching massive BFS queries [23,58].…”

C-SAW: A Framework for Graph Sampling and Random Walk on GPUs

“…Third, an extremely large graph, which drives the needs of graph sampling and random walk, usually goes beyond the size of GPU memory. While there exists an array of solutions for GPU-based large graph processing, namely, unified memory [26], topology-aware partition [27] and vertex-range based partitions [28], graph sampling and random walk algorithms, which require all the neighbors of a vertex to present in order to compute the selection probability, exhibit stringent requirement on the partitioning methods. In the meantime, the asynchronous and out-of-order nature of graph sampling and random walk provides some unique optimization opportunities for out-of-memory sampling, which are neither shared nor explored by traditional out-of-memory systems.…”