Leveraging System Behaviour Data to Improve the Process of Load Testing Large Scale Software Systems

从电子商务网站(如Amazon和eBay)到电信基础设施(如BlackBerry)的许多大型软件系统都必须支持对数千或数百万用户的并发访问。研究表明，这些系统的许多现场问题是由于它们无法扩展以满足用户需求，而不是功能缺陷。无法扩大规模会导致灾难性的失败和媒体的不利报道(例如，苹果MobileMe的拙劣发布)。为了确保这些系统的质量，除了常规的功能测试程序(例如，单元测试和集成测试)之外，负载测试是必需的测试程序。随着越来越多的服务(如苹果的iCloud和Google Drive)在云中为数百万甚至数十亿用户提供服务，负载测试变得越来越重要。**负载测试通常指的是评估负载下的系统行为的实践。典型的负载测试使用一个或多个同时向被测系统发送请求的负载生成器。在负载测试过程中，监视被测系统，并记录千兆字节或千兆字节的系统行为数据(例如，性能计数器和执行日志)。在支持问题诊断和远程问题解决的大规模系统中广泛可用的系统行为数据包含关于外部环境(例如，网络等待时间或请求的速率)以及系统的内部执行状态(例如，请求失败或请求队列的大小)的丰富信息。然而，由于此类数据的大小和复杂性，负载测试从业者目前仅以有限的方式使用它：主要用于特别的高级手动检查(例如，崩溃检查和内存泄漏检查)。使用这些有价值的信息来改进负载测试过程的软件测试研究很少。**本研究的长期目标是利用系统行为数据中包含的丰富信息来改进大型软件系统的负载测试过程。本研究旨在改进负载测试的所有三个阶段的理论和实践：测试设计、测试执行和测试分析。针对长期目标，提出了以下短期研究目标：(1)系统地验证负载测试套件(1名博士生)；(2)自适应负载测试执行(2名博士生)；(3)深入和可伸缩的负载测试分析(2名博士生)。预期的研究结果将有助于负载测试从业者和对大型软件系统的监控、测试和分析感兴趣的软件工程研究人员。

Many large scale software systems ranging from e-commerce websites (e.g., Amazon and Ebay) to telecommunication infrastructures (e.g., BlackBerry) must support concurrent access to thousands or millions of users. Studies show that many field problems of these systems are due to their inability to scale to meet user demands, rather than feature bugs. The inability to scale causes catastrophic failures and unfavorable media coverage (e.g., the botched launch of Apple's MobileMe). To ensure the quality of these systems, load testing is a required testing procedure in addition to conventional functional testing procedures (e.g., unit testing and integration testing). Load testing is gaining more importance, as an increasing number of services (e.g., Apple's iCloud and Google Drive) are being offered in the cloud to millions or even billions of users.**Load testing, in general, refers to the practice of assessing the system behavior under load. A typical load test uses one or more load generators that simultaneously send requests to the system under test. During the course of a load test, the system under test is monitored and gigabytes or terabytes of system behaviour data (e.g., performance counters and execution logs) is recorded. The system behaviour data, which is widely available in large scale systems to support problem diagnosis and remote issue resolution, contains rich information about the external environment (e.g., network latency or the rate of the requests) as well as the internal execution state of the system (e.g., request failures or the size of the request queues). However, due to the size and complexity of such data, load testing practitioners currently use it in a limited manner: mainly for ad-hoc high level manual checks (e.g., crash checks and memory leak checks). Little software testing research has been done to improve the load testing process with such valuable information. **The long term goal of this research is to leverage the rich information contained in the system behaviour data to improve the process of load testing large scale software systems. This proposed research aims to improve the theory and practices in all three phases of a load test: test design, test execution and test analysis. The following short term research objectives are proposed towards the long term goal: (1) Systematically Validating Load Test Suites (1 PhD student); (2) Adaptive Load Test Execution (2 Master's students); (3) In-depth and Scalable Load Test Analysis (2 PhD students). The expected research outcome will be useful for load testing practitioners and software engineering researchers with interest in monitoring, testing and analyzing large scale software systems.