Towards Accelerating Data Intensive Application's Shuffle Process Using SmartNICs

The wide adoption of the emerging SmartNIC technology creates new opportunities to offload application-level computation into networking layer, which frees burden host CPUs, leading performance improvement. Shuffle, all-to-all data exchange process, is a critical building block for network communication in distributed data-intensive applications and can potentially benefit from SmartNICs. In this paper, we develop SmartShuffle, accelerates application's shuffle process by offloading various tasks devices. SmartShuffle supports both low-level functions, including partitioning transport, high-level tasks, filtering, aggregation, sorting. adopts coordinated architecture make sender-side receiver-side SmartNICs jointly contribute benefits offload. carefully manages tight time-varying memory constraints on device. We propose liquid approach, dynamically migrates operators between CPU at runtime such that resources devices are fully utilized. prototype Stingray SoC plug it Spark. Our evaluation shows improves efficiency I/O with lower job completion time. outperforms Spark, Spark RDMA up 40% TPC-H.