Network Interface Cards as First-Class Citizens

Network performance has improved by over an order-ofmagnitude in the past decade. While, the overall system architecture itself has made modification attempts to match such growth, these changes are mostly “incremental enhancements, tweaks and adjustments” that try to keep the hardware control hierarchy away from the network communication path. However, we are rapidly reaching a stage where these adjustments are no longer sufficient to allow the network to realize its full capacity without being bottlenecked by the hardware control hierarchy in the system. Thus, rather than a “band-aid” fix to high-performancenetworking and I/O at the compute node, in this paper we re-visit a more radical approach that elevates the network interface card from a second-class citizen that resides out on an I/O interconnect to a first-class citizen that resides on the system bus. In this architecture, the network adapter would have its own cache, its own memory, its own processing units and follows the overall cache coherency and memory management protocols, much like what a regular CPU does. This architecture takes a step beyond existing system architectures and allows for direct communication data management without having to coordinate with the north bridge on every access.