Lightweight and Statistical Techniques for Petascale PetaScale Debugging

[摘要] This project investigated novel techniques for debugging scientific applications on petascale architectures. In particular, we developed lightweight tools that narrow the problem space when bugs are encountered. We also developed techniques that either limit the number of tasks and the code regions to which a developer must apply a traditional debugger or that apply statistical techniques to provide direct suggestions of the location and type of error. We extend previous work on the Stack Trace Analysis Tool (STAT), that has already demonstrated scalability to over one hundred thousand MPI tasks. We also extended statistical techniques developed to isolate programming errors in widely used sequential or threaded applications in the Cooperative Bug Isolation (CBI) project to large scale parallel applications. Overall, our research substantially improved productivity on petascale platforms through a tool set for debugging that complements existing commercial tools. Previously, Office Of Science application developers relied either on primitive manual debugging techniques based on printf or they use tools, such as TotalView, that do not scale beyond a few thousand processors. However, bugs often arise at scale and substantial effort and computation cycles are wasted in either reproducing the problem in a smaller run that can be analyzed with the traditional tools or in repeated runs at scale that use the primitive techniques. New techniques that work at scale and automate the process of identifying the root cause of errors were needed. These techniques significantly reduced the time spent debugging petascale applications, thus leading to a greater overall amount of time for application scientists to pursue the scientific objectives for which the systems are purchased. We developed a new paradigm for debugging at scale: techniques that reduced the debugging scenario to a scale suitable for traditional debuggers, e.g., by narrowing the search for the root-cause analysis to a small set of nodes or by identifying equivalence classes of nodes and sampling our debug targets from them. We implemented these techniques as lightweight tools that efficiently work on the full scale of the target machine. We explored four lightweight debugging refinements: generic classification parameters, such as stack traces, application-specific classification parameters, such as global variables, statistical data acquisition techniques and machine learning based approaches to perform root cause analysis. Work done under this project can be divided into two categories, new algorithms and techniques for scalable debugging, and foundation infrastructure work on our MRNet multicast-reduction framework for scalability, and Dyninst binary analysis and instrumentation toolkits.

[发布日期] 2014-06-30

[学科分类] 数学（综合）