Making In-Network Computing a Shared Infrastructure in Datacenters: Architecture, Challenges and Open Issues

Emerging programmable network devices spawn the trend of offloading computations into the network. Nowadays, in-network computing (INC) has been proven to be an essential way to break the performance bottleneck of distributed applications in datacenters. Various applications have achieved orders of magnitude performance improvement, such as distributed machine learning, consensus, and caching, via INC. Currently, optimizations of specific applications with INC are widely discussed, but the widespread use of INC is challenging due to the lack of architecture for deploying multiple applications on a shared INC infrastructure. These problems narrow the generality and practicability of INC when deployed in the real environment. In this article, we analyze the requirements and challenges of the practical INC deployment in data-centers, and present our key insights. We propose the shared in-network computing network (SINCN) architecture to enable the practical deployment of multiple applications in a shared INC infrastructure, which aims to extend the use of INC and benefit more applications from INC. In SINCN, the INC primitives, programming abstractions, and runtime programmable technology are jointly leveraged to solve the problems of generality, efficiency, flexibility, and isolation. In addition, we propose some technical challenges that occur in real deployment and present some potential open issues. Finally, a typical use case with SINCN is introduced, and the simulation results show the advantages of SINCN in terms of resource efficiency.