大规模分布式 GPU 图嵌入在腾讯的实践之路.pdf

1、大规模分布式GPU图嵌入在腾讯的实践之路微信数据中心王洋子豪#page#Outline Background InformationGPU Graph Embedding System Designc Performance Analysiss Conclusion and Future Works#page#Background#page#Background Information: Graph DataSocial NetworksBiology NetworksFinance NetworksInternet of ThingsInformation NetworksLogistic

2、NetworksloT点#page#Background Information: EmbeddingReflect networkMaintain networkstructurepropertiesTransitivityTransform network nodes into vectors that are fit foroff-the-shelf machine learning models.点Reference: http；#page#Background Information: EmbeddingG=（VE）G=（V）VectorSpacegenerateembed点5%20

3、0n%620NE96Reference：HOALT7O#page#Background Information: ChallengesHow to design a parallel training method to enable scalable training ondistributed clusters？How to optimize memory-bound operations with uniqueirregularmemory access patterns to gain high performanceHow to design communication strate

4、gy to optimize inter-GPU CPU-GPUand inter-node communication？ Embedding task combines network-oriented primitives and machinelearning oriented primitives. How to make them collaborate efficiently？三#page#Background Information: ChallengesHow to design a parallel training method to enable scalable tra

5、ining onHow to optimize memory-bound operations with unique irregularmemory access patterns to gain high performance？Hierarchical Data Partitioning StrategyHow to design communication strategy to optimize inter-GPU CPU-GPUand inter-node communication？Pipelined Training Strategy Embedding task combin

6、es network-oriented primitives and machinelearning oriented primitives. How to make them collaborate efficiently？意Decouple graph-primitive and ML-primitive#page#SystemDesign#page#GPU Embedding System Design无在元子无子无子活天无子无活地GPU OGPU 1GPU 2GPUNode 0GPUGPUGPU 6GPU 7GPU OGPU 1GPU 2GPU3Node 1GPU 41GPU5GPU