Accelerating Inter-Host Communication between Microservices with RDMA/eBPF

As the cloud-native paradigm increases in popularity and quickly becomes the de-facto standard for datacenter software development, monolithic applications are split into functionally distinct microservices. Consequently, east-west network traffic becomes critical to ensure system correctness. Traditionally, node-to-node communication relies on the TCP/IP stack, which provides reliable, ordered message delivery between applications. Recently, technologies such as Remote Direct Memory Access (RDMA) have emerged, offering equivalent reliability while supporting higher data rates and lower latency. This thesis presents a TCP-to-RDMA proxy that bridges these technologies. The proxy transparently intercepts outbound application traffic at the socket level and leverages RDMA to transport it reliably, effectively bypassing the TCP/IP stack while maintaining full application compatibility. Achieving full transparency without kernel modifications is challenging; our solution exploits the eBPF framework, which enables in-kernel programmability and allows network traffic to be dynamically monitored and redirected. This approach ensures portability, low overhead, and seamless integration with existing applications. In addition to outlining the overall architecture, this work discusses the main design choices and the implementation of the proxy, focusing on how eBPF was leveraged to ensure transparency and flexibility, and how RDMA was exploited to achieve higher throughput. We also conducted a preliminary performance evaluation: although the results are not yet fully satisfactory, they already provide useful insights into the system’s behavior and point to clear directions for improvement. Along the way, we encountered unexpected issues with the SKMSG hook. While this was not part of the initial objectives, it turned out to be an interesting and valuable finding that sheds light on the limitations of current kernel mechanisms. Looking ahead, we believe that this line of work opens the door for a new generation of datacenter networking. By bridging legacy application interfaces with modern transport technologies in a fully transparent way, it becomes possible to rethink the role of the operating system in distributed environments, reduce the reliance on traditional network stacks, and unlock new opportunities for performance and scalability.