2026 P4 Workshop

Data-Plane Programmability Across Heterogeneous Networks

Fast, Flexible, and Intelligent AI-Native Networks

Our networks are becoming increasingly heterogeneous, from diverse programmable devices that implement scale-out Ethernet to the new protocols and communication semantics emerging in scale-up systems. In this talk, I will discuss how to enable data-plane programmability across this spectrum. I will first present Alkali, a compiler framework for programming heterogeneous SmartNIC architectures in scale-out networks. I will then discuss our work on extending programmability to memory-semantic scale-up interconnects. Together, these efforts explore how data-plane programming must evolve to accommodate both heterogeneous devices and emerging network protocols.

Next-generation networks increasingly need to make sophisticated, data-driven decisions at line rate, yet today’s solutions are typically either fast but inflexible or intelligent but slow. In this talk, I will present my research on bridging this gap by bringing machine learning directly into the network data plane. I will highlight Taurus, a programmable architecture for executing ML models at line rate, and Homunculus, a declarative framework that automatically translates high-level operator objectives into efficient in-network ML policies. I will conclude with my broader vision for building fast, flexible, and intelligent next-generation networks and systems.

• AC Hotel Sunnyvale Moffett Park
  • $536 avg/night 
  • 0.5 mile from event venue
• TETRA Hotel, Autograph Collection
  • $478 avg/night
  • 0.5 mile from event venue 
• Extended Stay America 
  • $207 avg/night
  • 1.0 mile from event venue 

• Radisson Hotel Sunnyvale
  • $275 avg/night
  • 1.3 mile from event venue 
• Hyatt Regency Santa Clara
  • $369 avg/night
  • 3.0 miles from event venue