Application Defined Networks

Abstract

Cloud applications today are no longer monolithic programs in the classical sense. They are graphs of microservices in which a single user request fans out into tens or hundreds of remote procedure calls (RPCs), and end-to-end latency and operating cost are determined not only by what each service computes but also by how those services communicate. Yet the way they communicate remains rooted in the Internet architecture. A typical RPC traverses a stack of general-purpose protocols: Protobuf, gRPC, HTTP/2, TLS, TCP, IP, while network policies are often enforced by an out-of-process proxy that must terminate that stack, parse it, and re-originate the request. The result is what this dissertation calls the RPC tax: a compound penalty in latency, CPU, and engineering complexity paid on every microservice call, in a setting, a single administrative domain, where much of the layered generality being paid for is not actually needed. This dissertation argues that the RPC tax is not inherent but a design choice, and it proposes a different approach: Application Defined Networks (ADN), a model in which the entire application network stack, serialization, end-to-end delivery, and application network functions, is specified once at a high level and compiled into a flat, application-specific implementation, rather than assembled at runtime from general-purpose layers. The thesis is organized around three pillars, each with its own complete artifact. First, MeshInsight establishes the problem empirically through a compositional cost model that decomposes an application network proxy into its component operations and shows that today's stack inflates application latency by up to 269% and CPU consumption by up to 163%, with protocol parsing as the dominant cost. Second, AppNet introduces a simple and analyzable model for application network functions together with a symbolic-equivalence engine that lets the compiler fuse, reorder, and reposition stateful elements without breaking observable behavior. Common application network functions, and production designs such as Meta's ServiceRouter and Google's Prequal, can be expressed in tens of lines of AppNet code, while compiler-driven optimization reduces latency by up to 82% and CPU by up to 75% over unoptimized baselines. Third, fRPC eliminates the layered stack itself: from a high-level specification, it synthesizes a policy-aware split-partition wire format, a non-terminating and non-buffering proxy, and shims that automatically resolve interference between delivery semantics and policy semantics, cutting end-to-end application latency by up to 52% and CPU usage by up to 33% versus the common deployment of gRPC + Envoy while preserving the same semantics and security guarantees. Together, the three artifacts demonstrate a compilation approach and a new way to build the application network substrate of the cloud: not as standardized infrastructure configured after deployment, but as an application-specific artifact generated from a specification.