Performance of Internet Applications on Modern Servers

平板电脑和智能手机等低成本、功能丰富的消费计算设备的出现，加上一系列多样化的基于互联网的“服务”，导致了大量的计算机服务器和网络基础设施。该基础设施的建设和运营成本都很高。在20世纪90年代中期，工业界和学术界都投入了大量的精力来分析互联网的性能，以提高性能并提供更具成本效益的解决方案。然而，自20世纪90年代末以来，功能等话题受到的关注远远超过了性能，因为许多互联网应用程序的横向可伸缩性使得在性能或可伸缩性问题上“投入更多硬件”成为事实上的解决方案。 理解和提高互联网应用程序性能再次成为一个重要问题，这是因为人们意识到，“扔更多硬件”的方法不仅增加了资本和运营费用，还增加了IT基础设施的环境足迹。此外，自从性能评估最后成为一个广泛流行的话题以来，计算机服务器的设计已经发生了很大的变化。例如，物理限制阻碍了处理器制造商继续提高时钟频率，以实现每个新一代处理器的合理性能提升。相反，增加处理器核心的数量已成为标准做法。虽然一些应用程序被(重新)设计为利用“多核服务器”，但其他应用程序现在才得到增强。此外，有关具有单核处理器的服务器的行为的知识仍然是用于操作现代多核系统的“传统智慧”。这导致部署的服务器比支持给定负载所需的服务器更多(从而增加了资本支出)。由于服务器在使用率较低时使用的电量高得不成比例，这增加了运营成本和IT的环境足迹。 拟议的研究是为了确定和演示如何充分利用现代服务器的能力，以便以最小的财务和环境成本满意地满足消费者的请求。这一结果可能会直接让贝尔和Telus等加拿大服务提供商受益。该计划将涉及三个研究线索： 1.Web服务器在现代计算机服务器上的实验评估。(1名博士生) 2.对现代计算机服务器上的入侵检测系统进行了实验评估。(1名博士生) 3.互联网流量的表征和建模。(2名硕士研究生) 主题#1和#2将探讨如何最大限度地利用多核服务器来应对现实的互联网工作负载。合成生成的工作负载或重放跟踪将在被测试的服务器上放置受控数量的工作负载。服务器的广泛工具将提供所需的数据，以便就如何配置、管理和操作因特网应用程序建立新的“传统智慧”，以便在最大限度地利用服务器的同时获得所需的服务质量。最初的研究(主题#1)将重点放在Web服务器上，因为影响这些服务器在实践中如何运行的能力可能会产生重大影响。主题2与互联网安全应用程序有关，预计这些应用程序将具有相当长期的重要性。卡尔加里大学现有的互联网监控基础设施由申请者运营，这激发了这一主题。线程#3将与其他线程并行发生。通过收集、分析和建模实际互联网流量，可以提高线程#1和#2所需工作负载的真实性，并增加研究的影响。

The emergence of low-cost, feature-rich consumer computing devices like tablets and smart phones, coupled with a diversifying set of Internet-based "services" has resulted in a substantial infrastructure of computer servers and networks. This infrastructure incurs large capital expenses to build and high operating expenses to run. In the mid-1990s, both industry and academia devoted considerable effort to analyze Internet performance, to both increase performance and to provide more cost-effective solutions. Yet since the late 1990s, topics like functionality have received much more attention than performance, as the horizontal scalability of many Internet applications made "throwing more hardware" at a performance or scalability issue the de facto solution. Understanding and improving Internet application performance has re-emerged as an important problem, due to a realization that the "throwing more hardware" approach not only increases capital and operating expenses, but also increases the environmental footprint of IT infrastructure. In addition, the design of computer servers has changed considerably since performance evaluation was last a widely popular topic. As an example, physical limitations have prevented processor manufacturers from continuing to increase the clock frequency to achieve reasonable performance gains with each new generation of processor. Instead, increasing the number of processor cores has become the standard practice. While some applications were (re)designed to take advantage of “multi-core servers”, others are only now being augmented. Furthermore, knowledge about the behavior of servers with single-core processors is still the "conventional wisdom" used to operate modern multi-core systems. This leads to the deployment of more servers (thus increasing capital expenses) than are needed to support a given load. As servers use a disproportionately high amount of power when lightly utilized, this increases operating costs and IT’s environmental footprint. The proposed research is to determine and demonstrate how to fully exploit the capabilities of modern servers, so that consumer requests can be satisfactorily served with a minimal financial and environmental cost. The results could directly benefit Canadian service providers like Bell and Telus. This program will involve three research threads: 1. Experimental evaluation of Web servers on modern computer servers. (1 PhD student) 2. Experimental evaluation of an Intrusion Detection System (IDS) on modern computer servers. (1 PhD student) 3. Characterization and modeling of Internet traffic. (2 Masters students) Threads #1 and #2 will explore how to maximize the use of multi-core servers for realistic Internet workloads. Synthetically generated workloads, or replayed traces will place controlled amounts of workload on the servers being tested. Extensive instrumentation of the servers will provide the data needed to establish new “conventional wisdom” on how to configure, manage and operate Internet applications, such that a desired Quality of Service can be obtained while maximizing the use of the servers. The initial research (Thread #1) will focus on Web servers, since the ability to influence how these are operated in practice could have a significant impact. Thread #2 relates to Internet security applications, which are expected to be of considerable long term importance. The Internet monitoring infrastructure in place at the University of Calgary and operated by the applicant motivates this thread. Thread #3 will occur in parallel to the other threads. By collecting, analyzing and modeling actual Internet traffic, the realism of the workloads needed in Threads #1 and #2 can be improved and the impact of the research increased.