Nexus: Transparent I/O Offloading for High-Density Serverless Computing

A research team led by JooYoung Park has introduced Nexus, a novel hypervisor designed specifically for high-density serverless computing. Current serverless architectures tightly couple application logic with I/O processing, forcing each virtual machine to duplicate heavy communication infrastructure like cloud SDKs, RPC, and TCP/IP stacks. This duplication consumes over 25% of a function's memory footprint and can double CPU cycles compared to bare-metal execution. While previous solutions like WebAssembly or library OSes attempted to address this, they sacrificed ecosystem compatibility, requiring developers to migrate code to new languages.

Nexus takes a fundamentally different approach by transparently decoupling compute from I/O at the hypervisor level. Built on KVM, it intercepts communication fabric at the API boundary and offloads it to an always-on host shared backend via zero-copy shared memory. This structural separation enables asynchronous I/O optimizations, including overlapping input payload prefetching with VM restoration from snapshots and writing output payloads back to storage off the critical path. The results are substantial: compared to production baselines, Nexus reduces overall node-level CPU consumption by up to 44% and memory consumption by 31%, increasing deployment density by 37%. It also reduces warm-start latency by 39% and cold-start latency by 10%, bringing response times within 20% of WASM-based alternatives while maintaining full ecosystem compatibility.

The system represents a significant advancement in serverless infrastructure efficiency. By preserving the conventional serverless programming model while dramatically improving resource utilization, Nexus enables cloud providers to host more functions on the same hardware without requiring developers to rewrite their applications. This breakthrough could lead to lower costs for serverless computing users and more sustainable cloud infrastructure overall, as fewer physical servers would be needed to handle the same workload.