I-GCN: A Graph Convolutional Network Accelerator with Runtime Locality Enhancement through Islandization

Abstract

Graph Convolutional Networks (GCNs) have drawn tremendous attention in the past three years. Compared with other deep learning modalities, high-performance hardware acceleration of GCNs is as critical but even more challenging. The hurdles arise from the poor data locality and redundant computation due to the large size, high sparsity, and irregular non-zero distribution of real-world graphs.In this paper we propose a novel hardware accelerator for GCN inference, called I-GCN, that significantly improves data locality and reduces unnecessary computation. The mechanism is a new online graph restructuring algorithm we refer to as islandization. The proposed algorithm finds clusters of nodes with strong internal but weak external connections. The islandization process yields two major benefits. First, by processing islands rather than individual nodes, there is better on-chip data reuse and fewer off-chip memory accesses. Second, there is less redundant computation as aggregation for common/shared neighbors in an island can be reused. The parallel search, identification, and leverage of graph islands are all handled purely in hardware at runtime working in an incremental pipeline. This is done without any preprocessing of the graph data or adjustment of the GCN model structure. Experimental results show that I-GCN can significantly reduce off-chip accesses and prune 38% of aggregation operations, leading to performance speedups over CPUs, GPUs, the prior art GCN accelerators of 5549×, 403×, and 5.7× on average, respectively. CCS CONCEPTS• Computer systems organization → Neural networks; Parallel architectures; • Computing methodologies → Parallel algorithms.

FlowGNN: A Dataflow Architecture for Universal Graph Neural Network Inference via Multi-Queue Streaming

“…Boost-GCN [59] specifically optimizes GCN via sparsity analysis and graph partitioning. I-GCN [15] is the most recent GCN accelerator delivering the state-of-the-art performance, which uses an islandization approach to de-duplicate redundant GCN computations by merging nodes with shared neighbors.…”

“…Complementary to GNN acceleration, another category of related work is general graph processors, such as GraphH [10], Blogel [52], Giraph++ [42], ForeGraph [11], FabGraph [36], HitGraph [61], AccuGraph [55], and the most recent open- Because of edge embeddings, redundancy removal by node merging in I-GCN [15] is not applicable. (c) SOTA graph processor ThunderGP [8] with pipelined scatter and gather, followed by a graph-level apply.…”

“…Additionally, the existence of edge embeddings prevent some GNN optimization techniques from being applicable. For example, the SOTA GCN accelerator, I-GCN [15], proposes a redundancy removal approach by merging nodes with same neighbors to reduce computation, as shown in Fig. 1(b) by merging node a and b.…”

“…Second, novel GNN models are rapidly emerging while accelerator innovation is lagging behind. For instance, most state-of-the-art (SOTA) GNN accelerators are tailored for graph convolutional networks (GCNs) [13,15,59,60], which can be conveniently expressed as sparse matrix multiplications (SpMM). However, the majority of GNNs are not suitable for SpMM because of complicated operations such as edge embedding, attention, mixed neighborhood aggregation, etc.…”

“…Feature SOTA GNN Accelerators Ours (FlowGNN) Generic : GCN only [59,60] : Support a wide types of GNNs with message-passing : Do not support edge embeddings [14,15,30,54] : Support edge embeddings, attention, multi-aggregation, etc. : Rely on graph-specific property (e.g., sparsity) [59] : Agnostic to GNN models, datasets, and graph structures Real-time : Requires heavy graph preprocessing [60] or partitioning on CPU [59] : Zero preprocessing; directly takes in raw graphs in a streaming fashion and processes on-the-fly Open-source : Not available yet : Implementation open-sourced with executables GNN family that can be represented as sparse matrix-matrix multiplications (SpMM) GIN [50] [20] [15] GNN family with edge embeddings and transformations where SpMM does not apply GAT [43] [3] Anisotropic GNN family with sophisticated message functions PNA [9] A popular GNN family arbitrarily using multiple aggregation methods DGN [4] A state-of-the-art GNN with a directional flow at each node and guided aggregation VN [16] A nism. Table 2 summarizes the currently supported GNN models, each being representative of a large GNN family.…”