Content area
Modern cloud datacenters support a wide range of distributed and data-intensive applications, from machine learning training to large-scale analytics and microservices. However, the networking abstractions available to these applications remain largely unchanged, continuing to expose a low-level, byte-stream interface that lacks awareness of application semantics. This rigid separation between applications and the network prevents cross-layer optimizations, forcing applications to either remain oblivious to network conditions or implement inefficient, ad hoc workarounds.
To address this gap, this dissertation explores pushing networking abstractions up the stack, enabling closer integration between application communication semantics and network infrastructure. Specifically, it investigates how cloud applications can leverage richer, intentaware networking services rather than treating the network as an opaque transport layer. By doing so, we can bridge the divide between application logic and network behavior, improving efficiency, adaptability, and manageability in cloud environments.
We present two fully implemented systems and one forward-looking design proposal that exemplify this approach. NetHint enables applications to dynamically optimize data transfers by exposing real-time network hints, breaking the cloud’s black-box networking model. It improves collective communication throughput by up to 2.7ˆ by allowing applications to adapt to network conditions. mRPC rethinks Remote Procedure Call (RPC) handling by moving it into a managed OS service, eliminating redundant processing in per-application proxies and reducing RPC latency by up to 2.5ˆ. Finally, NUSE explores a more flexible, hybrid kernel-userspace networking architecture, demonstrating how the OS itself can evolve to support diverse application-specific network processing needs. While NetHint and mRPC focus on immediate performance gains, NUSE represents a broader vision for making networking extensible in future cloud platforms.
Together, these contributions validate the core dissertation statement: that a closer integration between application communication semantics and network infrastructure is essential to improving efficiency and flexibility in cloud datacenters. By breaking traditional abstraction barriers and enabling controlled information sharing across layers, we achieve significant gains in performance, manageability, and scalability, without sacrificing security or generality. By exposing richer network semantics and shifting key communication services into the system layer, we pave the way for more efficient, adaptable, and programmable cloud networking architectures.