ZOC: Elastic and Cost-Efficient Virtual SmartNIC Architecture for Cloud Physical Machines
Abstract
DPU-based SmartNICs are reshaping cloud infrastructure by offloading core I/O functions, enabling cloud physical machines that support both bare-metal and virtualized environments. However, large-scale deployments face practical challenges—including operational inconsistency across heterogeneous fleets, limited resource elasticity, and performance bottlenecks caused by slow-path processing. We present ZOC, a software-defined virtual SmartNIC architecture that transforms general-purpose servers equipped with commodity NICs into full-featured cloud physical machines. ZOC delivers compatibility with existing infrastructure and provides a user experience on par with DPU-based nodes—without requiring any specialized hardware. At its core, ZOC integrates a passthrough NIC and dynamically provisioned host resources into a dedicated service VM, forming an elastic and programmable acceleration platform. Built upon this foundation, it introduces an efficient cross-domain device abstraction that exposes standardized I/O devices to the host, enabling seamless access to cloud storage and VPC networks. Extensive evaluation shows that ZOC achieves superior cost efficiency, eliminates slow-path bottlenecks, and provides high maintainability. Deployed in a major cloud, ZOC serves diverse public and private cloud services as a production-ready complement to existing DPU-based solutions.