Xiangyu Li

Ph.D. Candidate in Computer Science

School of Computer Science

College of Computing

Georgia Institute of Technology

Date: Monday, April 26, 2021

Time: 9:00 AM - 11:00 AM (ET)

Location: https://bluejeans.com/648661869

Committee:

Dr. Alessandro Orso - Advisor, School of Computer Science, Georgia Institute of Technology

Dr. David Devecsery - School of Computer Science, Georgia Institute of Technology

Dr. Qirun Zhang - School of Computer Science, Georgia Institute of Technology

Dr. Spencer Rugaber - College of Computing, Georgia Institute of Technology

Dr. Marcelo d'Amorim - Department of Computer Science, Federal University of Pernambuco

Abstract:

Software debugging is an expensive activity that is responsible for a significant part of

software maintenance cost. In particular, locating faulty code (i.e., fault localization) is

one of the most challenging parts. In the past years, researchers have proposed many techniques

that aim at fully automating the task of fault localization. Although these techniques are

shown to be effective in reducing the amount of code developers need to inspect to locate

faults, there is growing evidence that they provide developers with limited help in realistic

debugging scenarios. I believe that a practical automated debugging technique should have

human developers at the center of the debugging process rather than trying to completely

replace them.

In this dissertation, I present three of my techniques that support developer-centric automated

debugging. First, I present ENLIGHTEN, an interactive, feedback-driven fault localization

technique. ENLIGHTEN supports and automates developers’ debugging workflow as follows.

It 1) uses traditional statistical fault localization (SFL) to formulate an initial hypothesis

of where the fault may be; 2) identifies a relevant subset of execution that can help support

or refute the formulated hypothesis; 3) presents the developer with a query about the identified

execution subset in the form of a correctness question about the input-output relation of the

partial execution; 4) refines its hypothesis of the fault by using the developer’s feedback; and

5) repeats these steps until the fault is found. Second, I discuss my work on improving the

accuracy of dynamic dependence analysis, which is a powerful tool for developers to investigate

program behavior in an interactive debugging setting and a foundation for many automated debugging

techniques to model dynamic execution semantics. I present my finding that existing dynamic

dependence analysis techniques could miss the cause-effect relations between faults and the

observed failures if the faulty program states propagate via incorrect computation of memory

addresses. To address this limitation, I define the concept of potential memory-address dependence,

which explicitly represents this type of causal relations, and describe an algorithm that computes

it. Third, I present TESSERACT, a technique that improves the scalability of dynamic dependency

analysis in the context of interactive debugging. Many existing dependency-based debugging

techniques are shown to work well on short executions, but fail to scale to even modest-length ones.

TESSERACT addresses this limitation by utilizing a record-and-replay system to efficiently recreate

the failing execution, break it down into small time slices, and analyze these slices in a

parallelized, and on-demand fashion.