面向确定性云服务性能的智能多资源协同编排及优化研究

云计算环境下，不确定行为的来源复杂性和传播放大性为服务性能保障带来了巨大挑战。尤其面对由众多存在依赖关系的组件组成的复杂应用，现有云资源分配和调度方法存在计算失序、传输失调、组件失协、应对失策问题，使得应用端到端的性能难以保障。本项目选择从“智能感知”和“多资源协同”的角度，在干扰可控的计算、传输协调的通信、端到端的协同编排和运行时应对四个方面提出了一套完整的多资源分配方案，以期大幅度提高云应用的端到端性能确定性。其基本思想可概括为:智能干扰感知的计算和通信，多资源协同分配，自学习资源调整。即，提出基于机器学习聚类干扰预测方法提高干扰感知的场景覆盖能力，指导计算和通信资源分配；设计应用多阶段与系统多资源的时空置换协同分配方案，保障端到端规划性能；提出基于强化学习的方法解决运行时序列多步最优资源调整决策问题，保障运行时性能。研究成果有望大幅度提高现有云服务的性能确定性，推动云计算产业发展。

In cloud systems, the diversity and transmissibility of uncertainty behaviors bring great challenge to the performance predictability of cloud services. The problem becomes particularly severe for complex applications consisting of a large number of interdependent components or micro-services, due to their disordered computing, inharmonious transmission, uncoordinated components and bad runtime remedial actions. Considering the multi-stage, multi-path characteristics of complex applications, we adopt an intelligent multi-resource orchestration and reconfiguration method to improve the predictability of their end-to-end performance. Our proposal supports interference-aware computation resource allocation through a micro-benchmark and clustering-based interference prediction method, and supports harmonious network flow transmission by learning the evolution features of mix-flows. We further present a coordination mechanism to orchestrate the resource allocation and scheduling of components with respect to their space-time constraints. The runtime performance is then ensured through a reinforcement learning-based resource reconfiguration mechanism. It is expected that our proposal could significantly improve the performance predictability of cloud services, thereby promoting the development of cloud computing.