Rise of Big Data techniques has led to the requirement for low latency analysis of high-velocity continuous data streams in real time. Several solutions, including Stream Processing Systems (SPSs), have been developed to enable real-time distributed stream processing. However, emerging application scenarios such as smart cities and wearable assistance that involve highly variable data rates keep on posing new challenges to the established stream processing engines for maintaining cost-effective executions. To cater to such scenarios, many modern SPSs have been proposed that leverage Cloud environment. The run-time scalability incorporated in these SPSs is in their early adaptations and are based on fixed scale sizes. Moreover, these scaling approaches do not adequately consider both the structure of the hosted streaming applications and the characteristic features of the underlying Cloud environment. Achieving true cost benefits of orchestrating streaming applications on Cloud-based pay-as-you-go model while maintaining the desired QoS, necessitates that both these issues are accounted in making the scaling decisions. This work presents a heterogeneity-aware, efficient auto-scaling strategy StreamScale-H which addresses both these issues. Simulation experiments, on representative stream applications, indicate that the proposed StreamScale-H auto-scaling algorithm exhibits much better performance in comparison with the state-of-the-art algorithms.

大数据技术的兴起导致对实时连续高速数据流进行低延迟分析的需求。已经开发了包括流处理系统（SPS）在内的几种解决方案，以实现实时分布式流处理。但是，诸如智能城市和可穿戴辅助设备之类的新兴应用场景涉及高度可变的数据速率，继续给已建立的流处理引擎带来新的挑战，以维持具有成本效益的执行。为了适应这种情况，已经提出了许多利用云环境的现代SPS。这些SPS中包含的运行时可伸缩性是在其早期改编中的，并且基于固定规模。此外，这些扩展方法没有充分考虑托管流应用程序的结构和底层云环境的特征。在基于云的按需付费模型上实现流应用程序编排的同时，要实现真正的成本优势，同时保持所需的QoS，就必须在制定扩展决策时同时考虑这两个问题。这项工作提出了一种异构感知的，高效的自动缩放策略StreamScale-H，该策略解决了这两个问题。在代表性的流应用程序上的仿真实验表明，与最新算法相比，所提出的StreamScale-H自动缩放算法具有更好的性能。