Building Flexible Data Center Network Stacks for the Terabit Era

Abstract

Modern data center workloads demand end-host network stacks that sustain terabit-scale bandwidth alongside ??-scale latency, overwhelming traditional software TCP stacks with high CPU overheads. ASIC-based transport offloads deliver high performance and energy efficiency but sacrifice flexibility, hindering customization to diverse application and deployment needs. This thesis explores flexible stateful TCP offload using emerging programmable in-network accelerators. It tackles the core challenge of mapping TCP’s complex, stateful processing onto the restrictive programming models of resource-constrained hardware, enabling fine-grained data-path parallelization. We present FlexTOE and Laminar, two novel TCP stack offloads built on Network Processing Unit (NPU) and Reconfigurable Match-Action Table (RMT) architectures. Both eliminate all host TCP data-path CPU overheads, integrate transparently with existing applications, remain robust under realistic network dynamics, and crucially, retain software programmability. The design principles developed generalize beyond TCP and extend naturally to other accelerator architectures. Through extensive evaluation, we demonstrate that these practical designs achieve a meaningful balance of high-performance, energy efficiency, and flexibility, surpassing state-of-the-art software stacks and offering a viable, adaptable alternative to rigid hardware transports.