Research On Debugging Environments Of Parallel Programs In Cluster Systems

The process of debugging parallel programs is a very important phase in the cycle of software development. Parallel debugging is a complicated work. Research of debugging techniques of parallel programs has been made for many years. However, few of the problems have been successfully solved. In this dissertation, research has been made on the topics of clock of cluster system, time-cost of debugging, intrusion of interactive debugging operation, dynamic conditional breakpoint, and performance debugging issues. Some of the achievements are used in the implementation of a parallel debugging environment for cluster systems.The contribution of this dissertation arrives from the follows:1. The characteristics, methods and techniques of performance debugging are described. The framework of performance debugging usually comprises a part which is responsible for collecting the performance information and a part which analyses and displays the performance data. In allusion to the unification of techniques and interface of correctness and performance debugging, the concept of debugging environment, which means the integrated environment of correctness debugging tool and performance debugging tool, has been proposed.2. An event-based physical clock has been development. It integrated the advantages of physical clock and logical clock so that it can not only represent the time duration but also preserve the partial order of events. A postmortem clock adjustment algorithm has been designed. The correctness and astringency have also been proved The outstanding advantage of our algorithm is that there is not any extra communication overhead.The deterministic checkpoint based on state-freezing can successfully avoid nephew messages, domino effect and live lock. By3. means of recording transient message, all the four problems of checkpoint are solved. And the non-deterministic is settled by the technique of recording/replay. The primary benefit of this algorithm is simple and easy to implement.4. A new debugging mode, called pure mode, has been propsoed to decrease the intrusion of interactive debugging operations. The characteristics of intrusion are formally uncovered. The correctness of freezing and melting algorithms has also been proved.5. By using of dynamic instrumentation, we implement conditional breakpoint. The efficiency of this kind conditional breakpoint is much higher than traditional one. Compared with GDB, we are able to execute a program with conditional breakpoints from several times to hundreds times faster.6. Combining dynamic instrumenting techniques and state-freezing strategy, the interactive performance debugging mode based on state-freezing is proposed Dynamic instrumenting technique makes it possible to introduce interactive debugging mode in performance debugging. To reduce intrusion of interactive debugging operations, state-freezing strategy is certainly a good choice. Instrumentation at the line of source program makes the collection of performance data more accurate and flexible.7. A debugging environment for parallel programs in cluster systems, called DENNET, is implemented. DENNET plays a positive role in prompting the researches of debugging environment in cluster systems.