Probing and Monitoring live distributed many-core systems

这项研究的长期目标是通过提供构建高级探测和监控工具的算法和技术，支持复杂的在线多核分布式计算机系统的设计、调试、操作和维护。这些在线分布式系统在Web、IT和移动电话基础设施中广泛使用，但事实证明，调试和调优设计和操作极其困难和昂贵。因此，需要工具来精确地监视、测量和了解在现实生产负载下全速运行的分布式系统的行为和性能。事实上，当以较慢的速度和较小的负载运行时，可能无法检测到许多错误，因为这会改变系统的时序。 为了实现这一点，第一个短期目标是从执行堆栈中的所有层(系统级、用户级和虚拟机级)提取低开销但详细的执行信息，并使用所有可用的信息来源(例如，软件和硬件跟踪以及性能计数器)。第二个目标是开发和验证这些提取算法，以确保即使在具有50个或更多共享内存处理器(例如Intel MIC芯片)的较新架构上，也能很好地扩展并保持较低的开销。最后，第三个目标是开发一些专门的痕迹分析模块。这将确保拟议的算法和工具将收集所有可用的有用信息，并将以最有效的方式执行，即使在多核系统上也是如此，从而使计时数据保持有效。专门的跟踪分析模块将使用户能够快速了解系统行为。 高效、准确地监控分布式多核系统的能力将使加拿大高科技行业的系统设计师和操作员能够更快地调试和调整分布式应用程序，从而更高效地运行现有系统或更快地设计更好的新系统。

The long term objective of this research is to support the design, tuning, operation and maintenance of sophisticated online many-core distributed computer systems, by providing algorithms and techniques to build advanced probing and monitoring tools. These online distributed systems are used pervasively in the Web, IT and cellular phone infrastructure but prove extremely difficult and costly to debug and tune for design and operation. Thus, tools are needed to precisely monitor, measure and understand the behavior and performance of distributed systems running at full speed under realistic production loads. Indeed, many errors may not be detectable when running at slower speeds and smaller loads because this changes the system timing. To achieve this, the first short term objective is to extract low overhead but detailed execution information from all layers in the execution stack (system level, user-level and virtual machines level) and using all the available sources of information (e.g. software and hardware traces, and performance counters). A second objective is to develop and validate these extraction algorithms to insure that they scale well and maintain their low overhead even on the newer architectures with 50 or more shared memory processors (e.g. Intel MIC chips). Finally, a third objective is to develop a number of specialized trace analysis modules. This will insure that the proposed algorithms and tools will collect all the useful information available, and will do it most efficiently, even on many-core systems, such that the timing data remains valid. The specialized trace analysis modules will allow users to quickly understand the system behavior. The ability to efficiently and accurately monitor distributed many-core systems will enable system designers and operators in the Canadian high technology industry to more rapidly debug and tune distributed applications, thus operating existing systems more efficiently or designing better new systems faster.