Download presentation
Presentation is loading. Please wait.
Published byFay Sullivan Modified over 9 years ago
1
Author: Chao Liu, Xiangyu Zhang Presented by Wenbin Li
2
Automated failure reporting is widely used failure indexing: which asks how to identify all failures due to the same fault ◦ Failure Prioritization ◦ Duplicate Failure Removal ◦ Patch Suggestion
3
Crashing failure: ◦ Incurred by memory bugs ◦ Nearly one-to-one mapping Noncrashing failure: ◦ Caused by semantic bugs ◦ A hard problem
4
T-Proximity ◦ Weakness: failures with similar behaviors rather than the same fault are indexed together R-Proximity ◦ Weakness: Use passing execution, which may violate the privacy of users.
5
identify a subset of program statements that are involved in producing a program failure. execution instances: it is used to distinguish different execution of the same statement s. Suppose s is executed n times, we use s1; s2; …; sn to denote the n execution instances. Data slice and full slice will be used in this study.
8
For example: E i = (10, 30, 40) E j = (10, 20, 30, 40) D(e i, e j ) = 1- ¾ = 1/4
9
To prepare for the failure indexing, we should do the following: 1. “due to” function 2. partitions the set of failuresX into m mutually exclusive and collectively exhaustive sets 3. define G(x i ) includes all the failures due to the same fault as x i
10
The distance function D quantifies how failures are close to each other based on the similarity between their corresponding failure signature Use a pair-wise distance matrix M(Ǿ;D), which is called the proximity matrix, to calculated for the given set of n failures
11
Two Concepts: ◦ Cohesion: To what extent failures in the same group are close to each other; ◦ Separation: To what extent failures in different groups are separated from each other. Both Cohesion and Separation should be HIGH
13
Problem: objects reside in a very high- dimensional space. Each failure is in a space of hundreds of dimensions because a typical slice contains hundreds of statements. Method: multidimensional scaling (MDS) techniques. The techniques re-arrange them in a specific way in a much lower (usually 2) dimensional space.
17
The two faults are manually seeded. hand-crafted test inputs are used. Metrics may not be accurate enough
18
The approach is more fault-relevant than the method used by T-PROXIMITY. The approach eliminates the need of passing executions, which is used by R-PROXIMITY.
19
It is hard to define data slices, considering how many codes a software may contain. The paper just provides some special example, which is not enough to prove the effectiveness of the approach.
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.