In-Network Computing
Overview
With the slowdown of Moore’s law, large-scale data centers have been employing an increasing number of domain-specific accelerators (GPUs, TPUs, and FPGAs) to deliver the needed unprecedented performance for computation-intensive workloads like machine learning model training and inference. Under this trend, the emergence of programmable network devices (P4 switches, SmartNICs, and DPUs more recently) has motivated a new concept called in-network computing, where programmable network devices are instructed (with P4/NPL languages) to accelerate application-specific computations (e.g., AllReduce for distributed machine learning training) in addition to running network functions. In-network computing brings tremendous performance benefits for a variety of distributed workloads, but also imposes challenges to the design of data center systems. Our work on in-network computing centers around programming and management systems, aiming to facilitate the adoption in production data centers.

