确定性跨域接口的自动合成
Automated Synthesis of Deterministic Cross-Domain Interfaces
浏览论文内容
中文总结 AI 辅助
该研究提出框架，利用LLM实现智能体与处理程序模块，自动合成跨域确定性网络的静态、动态契约，经实验验证其动态契约性能优于静态配置，且分工模式可保障模型可靠性。
中文摘要 AI 辅助
确定性网络覆盖异构域，在每个边界处，两个域必须就假设-保证契约达成一致：客户端可注入的流量，以及运营商将为其维持的QoS。将此类契约组合为端到端保障是标准化的，但推导每个域的契约并非如此。如今这些契约是手工制作的、静态的，且过度配置。当流量变化且契约必须动态时，难度会上升。我们提出了一个框架，可自动合成静态和动态类别的每个域契约，以及动态契约所依赖的注册机制。该框架由两个使用大型语言模型（LLM）实现的推理模块驱动：智能体（agent）接收客户端的流量声明并搜索运营商的配置机制，处理程序（handler）从底层的类型化披露中构建网络模型。处理程序调用网络演算内核处理模型的硬性条款，以及独立的预言机——一个可靠的模拟器、测试床或实时网络——该预言机会验证每个候选方案，并发现哪些内容缺乏先验代数形式：何时重配置是安全的，以及应用重配置的时机。我们为TDM-PON上的上行5G前传合成了一个动态契约，基于数据包级模拟器进行验证。在来自两个LLM家族的六次尝试中，每次合成都产生了一个可行的、经验证的契约，其精度约为1.2μs，保持着100μs的截止时间，而反应式调度无法满足该截止时间，其带宽效率最高是静态过度配置的3.5倍。而当要求LLM直接计算最坏情况延迟时，六次尝试中有五次是不健全的——这证明了分工的合理性：LLM构建模型，形式工具掌握数值权威。同一框架无需修改，便在第二个底层上合成了静态5G-TSN桥接契约。
英文摘要
Deterministic networking spans heterogeneous domains. At each boundary, two domains must agree on an assume--guarantee contract: what traffic the client may inject, and the QoS the carrier will hold for it. Composing such contracts into an end-to-end guarantee is standardized, but deriving each domain's contract is not. Today they are hand-crafted, static, and over-provisioned. The difficulty rises when traffic changes and the contract must become dynamic. We present a framework that automatically synthesizes the per-domain contract for both static and dynamic classes, along with the registration the dynamic one rests on. Two reasoning modules implemented with large language models (LLMs) drive it: an agent takes the client's traffic declaration and searches the carrier's configuration mechanisms, and a handler builds the network model from a typed disclosure of the substrate. The handler calls a network-calculus kernel for the model's hard terms, and an independent oracle---a faithful simulator, testbed, or live network---which verifies each candidate and discovers what lacks an a-priori algebraic form: when a reconfiguration is safe, and the instant to apply it. We synthesized a dynamic contract for uplink 5G fronthaul over a TDM-PON, grounded against a packet-level simulator. Across six draws from two LLM families, every synthesis produced a feasible, verified contract tight to ${\sim}1.2\,μ$s, holding a $100$-$μ$s deadline that reactive scheduling cannot meet, at up to $3.5$ times the bandwidth efficiency of static over-provisioning. Tasked instead with computing the worst-case delay directly, the LLMs were unsound in five of six attempts---evidence for the division of labor: LLMs construct the model, formal tools hold numeric authority. The same framework, unchanged, synthesized a static 5G--TSN bridge contract on a second substrate.
