[Paper Review] AWB-GCN: A Graph Convolutional Network Accelerator with Runtime Workload Rebalancing

Tong Geng, Ang Li+9 more|arXiv (Cornell University)|Aug 23, 2019

Advanced Graph Neural Networks64 references9 citations

TL;DR

AWB-GCN is a hardware accelerator for Graph Convolutional Network (GCN) inference that uses runtime workload rebalancing via dynamic distribution smoothing, remote switching, and row remapping to address extreme workload imbalance in sparse, power-law distributed graphs. It achieves 3,255× speedup over CPUs, 80.3× over GPUs, and 5.1× over prior accelerators on FPGA evaluation with 4K processing elements.

ABSTRACT

Deep learning systems have been successfully applied to Euclidean data such as images, video, and audio. In many applications, however, information and their relationships are better expressed with graphs. Graph Convolutional Networks (GCNs) appear to be a promising approach to efficiently learn from graph data structures, having shown advantages in many critical applications. As with other deep learning modalities, hardware acceleration is critical. The challenge is that real-world graphs are often extremely large and unbalanced; this poses significant performance demands and design challenges. In this paper, we propose Autotuning-Workload-Balancing GCN (AWB-GCN) to accelerate GCN inference. To address the issue of workload imbalance in processing real-world graphs, three hardware-based autotuning techniques are proposed: dynamic distribution smoothing, remote switching, and row remapping. In particular, AWB-GCN continuously monitors the sparse graph pattern, dynamically adjusts the workload distribution among a large number of processing elements (up to 4K PEs), and, after converging, reuses the ideal configuration. Evaluation is performed using an Intel D5005 FPGA with five commonly-used datasets. Results show that 4K-PE AWB-GCN can significantly elevate PE utilization by 7.7x on average and demonstrate considerable performance speedups over CPUs (3255x), GPUs (80.3x), and a prior GCN accelerator (5.1x).

Motivation & Objective

• Address the critical performance bottleneck in GCN inference caused by extreme workload imbalance in real-world sparse graphs with power-law degree distributions.
• Overcome limitations of existing accelerators—such as SCNN and SIGMA—that rely on global element-wise controllers or are ill-suited for ultra-sparse, irregular SpMM workloads.
• Design a lightweight, scalable hardware autotuning framework that dynamically rebalances workloads with minimal area and delay overhead.
• Enable efficient utilization of large-scale processing element arrays (up to 4K PEs) in FPGA-based accelerators for GCN workloads without relying on FPGA-specific features.
• Provide a general-purpose, extensible accelerator architecture applicable to other GNNs relying on sparse matrix multiplication (SpMM), such as GraphSAGE and GINConv.

Proposed method

• Proposes three runtime workload rebalancing techniques: distribution smoothing (local load balancing among neighboring PEs), remote switching (global row-level task migration), and row remapping (reassigning rows to balance PE utilization).
• Employs a coarse-grained, lightweight controller that monitors workload distribution across PEs in real time and triggers rebalancing only when imbalance exceeds a threshold.
• Uses dynamic distribution smoothing to redistribute non-zero pairs among nearby PEs to reduce hotspots, avoiding the high cost of global element-wise control.
• Implements remote switching and row remapping as global, row-wise controllers that can reassign entire rows or row segments to underutilized PEs, enabling large-scale load balancing.
• Integrates autotuning to learn and reuse optimal workload configurations after convergence, minimizing reconfiguration overhead.
• Designs the accelerator to handle ultra-sparse matrices (sparsity >99.9%) typical in GCNs by avoiding pre-scanning and minimizing index storage overhead.

Experimental results

Research questions

• RQ1 How can hardware-based autotuning effectively reduce workload imbalance in GCN inference on ultra-sparse, power-law distributed graphs?
• RQ2 What are the most efficient and scalable hardware techniques to rebalance SpMM workloads across thousands of PEs without incurring prohibitive area or latency costs?
• RQ3 Can a hybrid controller combining local and global rebalancing strategies outperform existing global element-wise controllers in terms of performance and resource efficiency?
• RQ4 To what extent can runtime workload rebalancing improve PE utilization and throughput in GCN accelerators compared to static or software-based load balancing?
• RQ5 How does the proposed architecture generalize across diverse GCN workloads and other SpMM-based GNNs such as GraphSAGE and GINConv?

Key findings

• AWB-GCN achieves an average 7.7× improvement in PE utilization across five real-world GCN datasets on an Intel D5005 FPGA.
• The 4K-PE version of AWB-GCN delivers 3,255× speedup over a high-end CPU, 80.3× over a GPU, and 5.1× over a prior GCN accelerator.
• Runtime workload rebalancing via distribution smoothing, remote switching, and row remapping enables sustained high utilization even on highly irregular, power-law graphs.
• The autotuning-based controller achieves performance gains with negligible area and delay overhead, avoiding the scalability issues of global element-wise controllers.
• The accelerator’s design is portable and not FPGA-specific, enabling deployment on other hardware platforms.
• The framework generalizes well to other GNNs relying on SpMM operations, such as GraphSAGE, GINConv, and GTN, due to shared computational primitives.