[Paper Review] Adaptive Work-Efficient Connected Components on the GPU

Michael Sutton, Tal Ben‐Nun+3 more|arXiv (Cornell University)|Dec 4, 2016

Parallel Computing and Optimization Techniques3 references3 citations

TL;DR

This paper presents an adaptive, work-efficient GPU implementation of the Connected Components algorithm using a modified Hook-Compress approach. By fusing Compress operations into a single Multi-Jump kernel and dynamically segmenting edges based on graph degree, it reduces atomic contention and CPU-GPU round trips, achieving up to 6.76× speedup over state-of-the-art GPU CC implementations across diverse graph types.

ABSTRACT

This report presents an adaptive work-efficient approach for implementing the Connected Components algorithm on GPUs. The results show a considerable increase in performance (up to 6.8$ imes$) over current state-of-the-art solutions.

Motivation & Objective

• To address the performance limitations of existing GPU-connected components algorithms due to high atomic contention and frequent CPU-GPU synchronization.
• To improve work-efficiency in parallel graph processing by minimizing redundant operations and reducing memory access overhead.
• To design a GPU-native, adaptive algorithm that dynamically adjusts to graph characteristics such as average degree and connectivity patterns.
• To eliminate the need for multiple sequential kernel launches by fusing Compress operations into a single, scalable kernel.

Proposed method

• Introduces a Multi-Jump kernel that fuses multiple Jump operations into a single parallel kernel, reducing kernel launch overhead and improving memory coalescing.
• Proposes an Atomic-Hook kernel that performs hooking with atomic compare-and-swap operations, ensuring correctness under concurrent updates without global synchronization.
• Employs a dynamic segmentation strategy that splits the edge list into approximately 2|E|/|V| segments based on average degree to balance atomic contention and Compress cost.
• Uses a partial-order scheduling in Multi-Jump to prioritize higher-indexed vertices (closer to roots), improving memory access patterns and reducing divergence.
• Applies a heuristic segmentation size of 2|E|/|V| to balance atomic operation cost and Compress overhead, adapting to graph density.
• Integrates the optimized kernels into a host-side loop that alternates between segmented Atomic-Hook and full Multi-Jump phases, enabling adaptive convergence.

Experimental results

Research questions

• RQ1 How can the work-efficiency of GPU-based Connected Components be improved by reducing atomic operations and kernel launch overhead?
• RQ2 What is the optimal number of edge-segments for balancing atomic contention and Compress cost in the Hook-Compress algorithm?
• RQ3 Can fusing multiple Compress operations into a single kernel significantly reduce performance bottlenecks in irregular graph workloads?
• RQ4 How does dynamic segmentation based on graph degree affect convergence speed and scalability across diverse graph types?
• RQ5 To what extent can device-centric kernel design reduce CPU-GPU round trips and improve overall performance in irregular graph algorithms?

Key findings

• The proposed adaptive algorithm achieves up to 6.76× speedup over the baseline Soman et al. [3] implementation on diverse graph workloads.
• The Multi-Jump optimization consistently improves performance across all graph types by reducing kernel launch overhead and improving memory access patterns.
• Atomic-Hook with dynamic segmentation yields the highest gains on low-degree graphs (e.g., 4.15× speedup on USA OSM), where atomic contention is most problematic.
• The optimal segmentation size of 2|E|/|V| consistently maximizes performance across all tested graphs, validating the heuristic design.
• For high-degree graphs like kron-logn21 and soc-live-journal, the Multi-Jump optimization has minimal impact (1.02× and 1.14×) due to the relatively small O(|V|) Compress cost compared to O(|E|) Hook cost.
• The adaptive approach reduces CPU-GPU round trips to a minimum by performing all work on the GPU, enhancing scalability and reducing latency.