Diagnosing Parallel and Distributed Software Performance Problems

点击翻译按钮获取中文摘要

Parallel and distributed computing will not enter the mainstream of computing until simple means are found to diagnose and correct performance problems. This award supports development of a system which, given monitored data collected from multiple program execution observations, will trace back from a graph of a performance metric to the processes and portions of source code that should be modified to improve the metric. The trace back is possible because all parts of the system, the source code, monitored data, graphs of performance metrics, and empirical and theoretical models, use a common state representation. The system performs three novel functions. First, it allows visual editing of monitored data from a small number of program executions to define aggregate states and an aggregate state transition rate matrix that can represent all observed program executions. Second, it generates graphs of performance metrics from the empirical model, rather than from the monitored data, to accurately predict discontinuities in program behavior. Third, it generates a theoretical model of program behavior that maps program behavior to a multidimensional Cartesian space. The system is applied to a TCP/IP protocol implementation and a parallel discrete event simulation system. Success in diagnosing performance problems in these large systems could change the way programmers design parallel and distributed software.