Yan, LeiPan, YueyangZhou, DiyuCandea, GeorgeKashyap, Sanidhya2025-06-092025-06-092025-06-062024-04-2210.1145/3627703.3629591https://infoscience.epfl.ch/handle/20.500.14299/251163WOS:001482612800067This paper presents NFOS, a programming model, runtime, and profiler for productively developing software network functions (NFs) that scale on multicore machines. Writing shared-state concurrent systems that are both correct and scalable is still a serious challenge, which is why NFOS insulates developers from writing concurrent code. In the NFOS programming model, developers write their NF as a sequential program, concerning themselves with the NF logic instead of parallelism and shared-state synchronization. The NFOS abstractions are both familiar to the NF programmer and convey to the NFOS runtime crucial information that enables it to correctly execute the NF's packet processing in parallel on multiple cores. Paired with NFOS's domain-specific concurrent data structures, this parallelism scales the NF transparently, obviating the need for developers to write concurrent code. We show that serial, stateful NFs run atop NFOS achieve scalability on par with their concurrent, hand-optimized counterparts in Cisco VPP [8]. Some scalability bottlenecks are inherent to the NF's semantics, and thus cannot be resolved while preserving those semantics. NFOS identifies the root causes of such bottlenecks and provides scalability recipes that guide developers in relaxing the NF's semantics to eliminate these bottlenecks. We present examples where such NFOS-guided relaxation of NF semantics further improves scalability by 2x to 91x.EnglishNetwork functionsTransparent scalingConcurrencyPerformance profiling and debuggingtext::conference output::conference proceedings::conference paper