Serverless In-Network Acceleration with FPGAs (SINA)

• Applicants: Professor Dr. Marco Platzner; Professor Lin Wang, Ph.D.
• Subject Area: Computer Architecture, Embedded and Massively Parallel Systems ; Security and Dependability, Operating-, Communication- and Distributed Systems
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 566724436

Data center computing is crucial for online services, but challenges arise from Moore’s law slowdown and surging data transfer. Innovations like in-network computing, which speeds up computations during data transfer, have shown significant performance gains. However, adoption is hindered by a lack of flexibility in both development and deployment: (1) Developers must customize services for specific network setups and device constraints. (2) Updating deployed services disrupts critical network operations. Additionally, data center operators are reluctant to expose low-level network details and resources due to complexity and security concerns. The main research goal of project SINA (serverless In-Network Acceleration with FPGAs) is to make in-network computing a first-class citizen for data center computing by enabling serverless in-network acceleration with FPGAs. This approach uses FPGAs integrated with programmable network devices as general accelerators, overcoming the limitations of current network devices and allowing flexible development of in-network services. Bringing the serverless model into in-network computing decouples services from the network infrastructure, facilitating flexible deployment. This combination reassures data center operators since core network components remain unexposed. With SINA, distributed applications can offload heavy computations to network-integrated accelerators as serverless functions on FPGAs, boosting performance and efficiency. SINA will achieve its goal by making the following contributions. First, we will develop a serverless runtime for network-integrated FPGAs based on our previous work ReconOS and extend it to support function compositions coordinated with programmable switches. Then, we will design abstractions and frameworks with a new programming model based on our previous work NetCL, a two-level resource management framework, and mechanisms for reliability and fault tolerance. Finally, we will build a prototype and implement several popular in-network services as well as SINA-enabled applications as case studies. SINA will spearhead the next generation of data center computing, showcasing the potential of hardware-software co-design for integrated computing and communication. By making in-network computing more flexible, SINA will help boost the performance and efficiency of millions of data center services powering our digital society.

• DFG Programme: Research Grants