PRISMA: Efficient Algorithms and Methods for Online Extraction of Performance Models in Virtualized Environments

现代软件系统正变得越来越复杂和动态。系统通常具有分层的体系结构，其中包括应用程序组件、中间件平台、虚拟机、管理程序和物理硬件。每一层都会影响系统的性能，然而，在明确考虑配置和部署参数的同时，如何分离、隔离和量化每一层对性能的影响是一项重大挑战，这些参数在运行过程中可能会动态变化。在过去的五年中，分层的体系结构级性能模型作为运行时性能管理的强大工具变得越来越流行，然而，这种模型通常手动构建成本很高，并且需要在受控环境中进行大量的实验分析。PRISMA项目的目标是开发新的方法，在系统运行期间自动提取虚拟化平台及其托管应用程序的架构级性能模型。该项目将为具有集成模型提取功能的虚拟化平台和虚拟设备提供一种新颖的参考架构。提取将完全基于集成到虚拟化和中间件级别的平台中的通用模型骨架（可组合模型构建块），并监视在系统运行时收集的数据，而不假设源代码的可用性或进行静态代码分析的可能性。提取的性能模型将捕获托管应用程序及其执行环境（包括虚拟化平台本身）的性能相关方面。因此，不再需要为容量管理构建性能模型而进行广泛而昂贵的实验分析。所开发的方法将促进绩效模型的创建，并将为利用这些模型进行主动绩效和资源管理奠定基础。通过虚拟化平台提供的自动化和集成在线模型提取、优化和维护功能，PRISMA将为预测模型在现实系统中的实际应用突破奠定基础。采用基于模型的运行时管理技术可以在不牺牲应用程序性能保证的情况下避免过度供应系统资源，从而显著提高现代虚拟化服务基础设施的效率。

Modern software systems are becoming increasingly complex and dynamic. Systems typically have a layered architecture which includes application components, middleware platform(s), virtual machine(s), hypervisor, and physical hardware. Each layer influences the system performance, however, it is a major challenge to separate, isolate and quantify the performance influences of each layer while explicitly taking into account configuration and deployment parameters which can change dynamically during operation. Over the past five years, layered architecture-level performance models have become increasingly popular as a powerful tool for run-time performance management, however, such models are usually costly to build manually and require extensive experimental analysis in a controlled environment. The aim of the PRISMA project is to develop novel methods for automatic extraction of architecture-level performance models of virtualization platforms and their hosted applications during system operation. The project will provide a novel reference architecture for virtualization platforms and virtual appliances with integrated model extraction capabilities. The extraction will be based solely on generic model skeletons (composable model building-blocks) integrated into the platforms at the virtualization and middleware level, and monitoring data collected at system run-time, without assuming availability of source code or possibility to conduct static code analysis. The extracted performance models will capture the performance-relevant aspects of the hosted applications and their execution environment including the virtualization platform itself. Thus, extensive and costly experimental analysis to build performance models for capacity management will no longer be necessary. The developed methods will facilitate the creation of performance models and will lay the foundation for proactive performance and resource management by means of the models. By automating and integrating online model extraction, refinement and maintenance as features provided by the virtualization platform, PRISMA will set the basis for a breakthrough in the practical use of predictive models in real-life systems. The adoption of model-based run-time management techniques promises to significantly improve the efficiency of modern virtualized service infrastructures by avoiding to over-provision system resources without having to sacrifice application performance guarantees.