一种数据驱动的图神经网络推理数据流感知在线调度方法.pdf

1、ASP-DAC 202523 Jan 2025 Tokyo,JapanA Data-Driven Approach to Dataflow-Aware Online Scheduling for Graph Neural Network Inferences1Pol Puidgemont1,Enrico Russo2,Axel Wassington1,Abhijit Das1,Sergi Abadal1,Maurizio Palesi21 Universitat Politcnica de Catalunya,Spain2 University of Catania,ItalyASP-DAC

2、202523 Jan 2025 Tokyo,JapanOverview2Efficiently running Graph Neural Networks inferences in online,real-time,or streaming contexts requiresoptimal mapping and dataflow configurations,which must adapt dynamically to the changing input graph.ASP-DAC 202523 Jan 2025 Tokyo,JapanOverview3Determining the

3、best dataflow configuration poses some challengesEfficiently running Graph Neural Networks inferences in online,real-time,or streaming contexts requiresoptimal mapping and dataflow configurations,which must adapt dynamically to the changing input graph.ASP-DAC 202523 Jan 2025 Tokyo,JapanBackgroundGr

4、aph Neural NetworksGraph Neural Networks(GNNs)are a specialized type of neural network designed toprocess and analyze graph-structured data,which consists of nodes(data points)and edges(relationships between points).4From left to right.Citronella molecule.Adiacency matrix representation.Graph repres

5、entation.From https:/distill.pub/2021/gnn-intro/ASP-DAC 202523 Jan 2025 Tokyo,JapanBackgroundGraph Neural NetworksGNNs are increasingly employed in a wide range of applications:social networkanalysis,recommendersystems,frauddetection,transportnetworksandlogistics,smart energy grids,cybersecurity,dru

6、g discovery,etc.5ASP-DAC 202523 Jan 2025 Tokyo,JapanBackgroundGraph Neural NetworksA Graph Neural Network(GNN)is a parametrized transformation of graph attributesdesigned to preserve permutation invariance and graph symmetries.In our case,theinput graph is represented by an adjacency matrix and real