Understanding computer systems through online analysis of architecture-level events.

摘要

Modern processors execute billions of seemingly indistinguishable instructions every second, yet these instructions are in reality a complex tapestry of specific applications, operating systems, middleware, and system libraries. Though, at the lowest level, the execution of such programs seems like a series of random bit manipulations, in reality, these executions do preserve, and in fact carry-out, the semantics of the full system as specified by the software developers. I propose that efficient extraction and analysis of such semantic information at the hardware level is possible, and that inferences drawn from this knowledge can provide useful insights ranging from complex networked system-wide characteristics to finer application level behaviors.;However, realizing these application level semantics at the hardware level poses challenging research problems, especially when systems are composed of multiple services, running across multiple machines. Firstly, we need an efficient way to non-intrusively extract relevant bits from the processor - I propose the use of 3D Integration technology to tap signals out of a processor and transport them across to a place where these bits can be analyzed. Secondly, we need efficient data-stream analysis techniques to summarize useful information from among the billions of bits that flow in every second - I present Range Adaptive Profiling to pinpoint various program "hot-spots". Thirdly, the summarized information for the various pieces of system-wide computations now needs to be understood in relation to the application semantic. This calls for understanding the interaction between different pieces of code executing on the processors and the information owing in and out of the network - I present Data Flow Tomography, a tool to extract and understand application and network level semantics at the hardware level. The dissertation elaborates on these three aspects of understanding modern computer systems and explains the proposed solutions with example real-world scenarios.