Towards efficient allocation of graph convolutional networks on hybrid computation-in-memory architecture

Graph convolutional networks (GCNs) have been applied successfully in social networks and recommendation systems to analyze graph data. Unlike conventional neural networks, GCNs introduce an aggregation phase, which is both computation- and memory-intensive. This phase aggregates features from the neighboring vertices in the graph, which incurs significant amounts of irregular data and memory access. The emerging computation-in-memory (CIM) architecture presents a promising solution to alleviate the problem of irregular accesses and provide fast near-data processing for GCN applications by integrating both three-dimensional stacked CIM and general-purpose processing units in the system. This paper presents Graph-CIM, which exploits the hybrid CIM architecture to determine the allocation of GCN applications. Graph-CIM models the GCN application process as a directed acyclic graph (DAG) and allocates tasks on the hybrid CIM architecture. It achieves fine-grained graph partitioning to capture the irregular characteristics of the aggregation phase of GCN applications. We use a set of representative GCN models and standard graph datasets to evaluate the effectiveness of Graph-CIM. The experimental results show that Graph-CIM can significantly reduce the processing latency and data-movement overhead compared with the representative schemes.