Exploring Domain-Specific Architectures for Network Protocol Processing

Existing architectures for network protocol processing incur bottlenecks in communication-intensive cloud applications. Software processing competes for scarce CPU cycles with applications, fixed protocol offloads such as RDMA lack required flexibility, and SmartNICs are inefficient, expensive, or powerhungry. The design space for alternative architectures remains largely unexplored, and choosing a suitable architecture is complicated by rapidly evolving application and transport protocols. We draw parallels to machine learning acceleration and argue for a similar approach: configurable domain-specific architectures (DSAs) combined with a domain-specific high-level programming model decoupling applications from architectures. We propose Kugelblitz, an implementation and design exploration framework for network DSAs enabling flexible and efficient protocol offload. Kugelblitz comprises abstractions for separately specifying hardware configuration and protocols, a hardware RTL generator, and a compiler to generate runtime configurations to implement specific protocols on specific hardware configurations.