AggrBoost: A Load-Balanced Computing Engine for Boosting Aggregation in GCN Accelerators

The emergence of Graph Convolutional Network marks a significant breakthrough in machine learning techniques for non-Euclidean spaces. During the graph aggregation process, several challenges arise, including load imbalance, low data reuse rate, and high off-chip data traffic. This paper proposes a computing engine for aggregation phase in GCNs based on a multi-edge parallel aggregation strategy with source node partitioning. This work featuring local sharing of source node feature vectors through source node partitioning and an encoding method named CSB (Compressed Sparse Block), achieves efficient load balancing under high parallelism through the equitable distribution of edge streams. It addresses memory access conflicts during the feature update process of destination nodes by utilizing distributed storage of partial sums, and effectively eliminates the pipeline stalls within partitions by Index-Driven MAC (Multiply-Accumulate). The PE utilization can exceed 97% on several popular datasets, making it suitable for application in GCN hardware accelerators.