Research Initiation Award: Adaptive Tracing for Optimal Parallel and Sequential Program Debugging

将开发用于并行和顺序程序的最佳跟踪和重播调试的自适应策略。随着并行机器变得越来越大，处理器变得越来越快，越来越多的跟踪瓶颈阻碍了调试所必需的复杂的基于跟踪的工具的开发。我们缓解这一问题的目标是双重的。首先，我们将探讨为提供执行回放而必须跟踪的内容的理论下限。我们将研究需要跟踪哪些消息、共享内存引用或局部变量访问，以便在有限的时间内重播执行的任何部分。通过限定重放时间，我们为调试查询提供了固定的响应时间，即使对于大规模或长时间运行的程序也是如此。其次，我们将开发最优或接近最优的策略，以适应正在调试的执行。我们的算法做出运行时跟踪决策，以适应下一步的执行，并保持最佳的小跟踪。＊＊＊

Adaptive strategies for the optimal trace-and-replay debugging of parallel and sequential programs will be developed. As parallel machines become larger and processors become faster, the increasing tracing bottleneck is preventing the development of sophisticated trace-based tools necessary for debugging. Our objectives to alleviate this problem are twofold. First, we will explore the theoretical lower bounds on what must be traced to provide execution replay. We will investigate which messages, shared-memory references, or local variable accesses need be traced to replay any part of the execution in a bounded amount of time. By bounding replay time, we provide a fixed response time to debugging queries, even for large-scale or long-running programs. Second, we will develop optimal or near-optimal strategies that adapt to the execution being debugged. Our algorithms make run-time tracing decisions to adapt to what the execution will be doing next, and keep traces optimally small. ***