High performance high availability distributed system performance analysis

从搜索引擎（如google.com, yahoo.com）到地理信息系统（如maps.google.com）、电子商务系统以及手持音乐和视频播放器，社会越来越依赖于复杂的计算机系统来进行许多应用。由于数据量、同时用户的数量或实时限制，此类系统具有严格的性能要求。然而，要达到预期的性能水平，需要几个软件程序、操作系统和通常是几台联网的计算机之间进行适当的交互。提出的研究计划的目标是开发工具和技术，用于单处理器，多处理器和分布式系统的高精度和低干扰系统级性能分析。对于使用多处理器、具有动态变化时钟频率、多线程抢占式操作系统（例如具有抢占的Linux）和运行在管理程序（例如Xen）上的多个虚拟机的较新的计算机系统，这尤其具有挑战性。所建议的系统在操作系统和用户级别使用源代码插装，以便收集重要事件的详细和精确的跟踪。插装需要将对总体性能和事件排序（调度）的影响降到最低。将开发专用分析和查看模块，以研究不同子系统（内存、磁盘I/O、网络等）的性能。与工业合作伙伴的协作已经开始，面临着非常苛刻的应用程序。这些行业合作伙伴将提供需要复杂性能分析工具的现实生活示例应用，以及对所提出算法、技术和工具的效率、准确性和适用性的反馈。

The society increasingly relies on sophisticated computer systems for numerous applications from search engines (e.g. google.com, yahoo.com) to geographical information systems (e.g. maps.google.com), eCommerce systems and handheld music and video players. Such systems have stringent performance requirements either because of the volume of data, the number of simultaneous users or real-time  constraints. Yet, achieving the desired level of performance assumes  proper interaction between several software programs, the operating system and often several networked computers.The objective of the proposed research program is to develop tools and techniques for highly accurate and low disturbance system level performance analysis of uniprocessor, multiprocessor and distributed systems. This is particularly challenging with newer computer systems using multiple processors, with dynamically changing clock frequency, multi-threaded preemptive operating systems (e.g. Linux with preemption) and multiple virtual machines running over an hypervisor (e.g. Xen).The proposed system uses source code instrumentation at the operating system and user level in order to gather a detailed and precise trace of significant events. The instrumentation needs to have a minimal impact on both overall performance and on events ordering (scheduling). Special purpose analysis and viewing modules will be developed to study the performance of different subsystems (memory, disk I/O, networking...). Collaboration is already under way with industrial partners faced with very demanding applications. These industrial partners will provide both real life sample applications needing sophisticated performance analysis tools, and feedback on the efficiency, accuracy and applicability of the proposed algorithms, techiques and tools.