MINERVA: Performance Prediction of Microservice Applications with Black-box Container Orchestration

现代分布式系统现在通常基于微服务架构，这被认为是软件开发人员的最佳实践。现代微服务应用程序通常部署在容器编排框架（例如Kubernetes）管理的容器中，该框架提供了许多自适应功能，如自动伸缩、智能负载平衡、断路器和故障恢复。虽然这些特性会显著影响应用程序的性能，但从应用程序开发人员的角度来看，它们使用的内部机制通常被认为是一个黑箱。这使得理解、量化、比较和优化微服务应用程序和编排框架的性能属性变得具有挑战性。会出现以下问题：应该如何配置适应机制以优化性能和运营成本之间的权衡？系统需要多长时间才能适应突然负载峰值为当前负载的三倍？如果某个服务实例或节点发生故障，预期的恢复时间和性能下降是多少？现有的性能预测方法只能用于评估稳态性能，通常需要对可能的自适应和自适应规则进行显式建模。由于大量潜在的适应以及许多适应机制依赖于复杂的机器学习模型这一事实，这是不可行的。对于某些方法，自适应逻辑本身可能在操作过程中发展。据我们所知，目前还没有一种基于模型的性能预测方法既考虑稳态阶段又考虑瞬态阶段，同时又不需要对自适应逻辑进行显式建模。该项目的目标是为带有黑盒容器编排的微服务应用程序开发这样一种方法。它应该使应用程序开发人员和系统操作员能够回答上述与性能相关的问题。该项目将有三个主要目标：(1)建模形式化，以捕获与预测具有黑盒容器编排环境中的现代微服务应用程序的瞬态和稳态性能相关的系统方面；(2)高效且可扩展的算法，用于模拟使用所提出的建模形式化建模的系统的性能；(3)新颖的分析算法和工作流，用于分析和解释通过开发的建模和仿真方法获得的结果。此外，我们将设计和实施一个新的基准和指标，以标准化的方式比较基于所提出的模拟框架的适应机制。将开展一系列案例研究，对不同的适应机制进行基准测试，并单独验证在项目下开发的方法、模型和工具，以及验证拟议的总体方法及其端到端性能。

Modern distributed systems are nowadays typically based on microservice architectures, which are considered a best practice among software developers. Modern microservice applications are commonly deployed in containers managed by a container orchestration framework (e.g., Kubernetes), which provides many self-adaptation features such as autoscaling, intelligent load balancing, circuit breakers, and failure recovery. While such features significantly influence the application performance, the internal mechanisms they use are normally considered a black box from the perspective of application developers. This makes it challenging to understand, quantify, compare, and optimize the performance properties of microservice applications and orchestration frameworks. Questions such as the following arise: How should the adaptation mechanisms be configured to optimize the tradeoff between performance and operating costs? How long would the system take to adapt to a sudden load spike of three times the current load? What would be the expected time-to-recovery and performance degradation if a certain service instance or a node fails? Existing performance prediction approaches can only be used to evaluate the steady-state performance and they normally require the explicit modeling of possible adaptations and adaptation rules. This is infeasible due to the large number of potential adaptations and the fact that many adaptation mechanisms rely on complex machine learning models. With some approaches, the adaptation logic itself may evolve during operation. To the best of our knowledge, there is no model-based performance prediction approach that considers both steady-state and transient phases, while not requiring the explicit modeling of the adaptation logic. The goal of the project is to develop such an approach for microservice applications with black-box container orchestration. It should enable application developers and system operators to answer performance-related questions like the above. The project will target three main goals: (1) modeling formalisms to capture the system aspects relevant for predicting both transient and steady-state performance of modern microservice applications in environments with black-box container orchestration, (2) efficient and scalable algorithms for simulating the performance of a system modeled using the proposed modeling formalisms, and (3) novel analysis algorithms and workflows to analyze and interpret the results obtained through the developed modeling and simulation approach. Further, we will design and implement a novel benchmark as well as metrics for comparing adaptation mechanisms in a standardized manner based on the proposed simulation framework. A series of case studies will be conducted to benchmark different adaptation mechanisms and to individually validate the methods, models, and tools developed under the project as well as to validate the overall proposed approach and its end-to-end performance.