Download presentation
Presentation is loading. Please wait.
Published byDevyn Prout Modified over 9 years ago
1
1 Bug Isolation via Remote Program Sampling Ben LiblitAlex Aiken Alice X. ZhengMichael Jordan Presented By : Arpita Gandhi
2
2 Motivation All deployed software systems have bugs Resources for improvement are limited The user community outnumbers the testing team Current systems are not systematic and automated
3
3 Related Applications Commercial databases produce log files and inspect them when the user reports a problem Netscape/Mozilla, Microsoft, GNOME,KDE have opt-in crash reporting systems Problems?
4
4 Goals Gather information systematically from user executions Perform automatic analysis to fix software quality problems using this information Speed up the process of bug isolation Develop an efficient, low overhead system
5
5 Discussion What do you think about the idea of using users as debuggers? Bugzilla showed 36,937 bugs and an additional 60,191 bugs as duplicates of already reported bugs
6
6 Design Goals Modest impact on program performance Statistically fair and uniformly random sampling Efficient transmission of client data
7
7 Sampling Blocks Consider the following piece of code { check(p != NULL); p = p->next; check(i < max); total += sizes[i]; } We want to sample 1/100 th of these checks
8
8 Sampling Blocks Solution 1 : Maintain a global counter modulo 100 Problem? for(i = 0 ; i < 10; i++) { check(p != NULL); p = p->next; check(i < max); total += sizes[i]; }
9
9 Sampling Blocks Solution 2: Use a random number generator { if(rnd(100) == 0)check (p != NULL); p = p->next; if(rnd(100) == 0) check (i < max); total += sizes[i]; } Problem?
10
10 Sampling Blocks Solution 3: Anticipate future samples Requires two versions of code: Slow path: Code with the sampled instrumentation Fast path: Code w/o the sampled instrumentation
11
11 Sampling Blocks Fast Path Code Slow Path Code if (countdown > 2) { p = p->next; total += sizes[i]; } if( countdown-- == 0 ) { check(p != NULL); countdown = getNextCountdown(); } p = p->next; if( countdown-- == 0 ) { check( i < max ); countdown = getNextCountdown(); } total += sizes[i];
12
12 Sampling Functions Represent sampling blocks as a CFG Weight of path is the maximum number of instrumentation sites Place a countdown threshold check on each acyclic region For each region r: If (next-sample countdown >weight) no samples taken
13
13 Instrumented Code Layout >4?
14
14 Optimizations Problem: A new threshold check before each function call Solution: Identify and ignore weightless functions No threshold checks, instrumented code Compiled without any modification
15
15 Optimizations Problem: Use of global variable for countdown management expensive Solution: Use Local Variables Local 1 Local 2 Local 3 global Function entry Use for threshold check, sampling decisions, decrements Function exit
16
16 Issues in Program Sampling Factors that hamper performance Remote monitoring Local storage Use of network bandwidth Storage on central server For smaller level deployment? Adaptive refinement of instrumentation
17
17 Other Issues Social and ethical concern Users’ interest in contribution Learning curve with software More testers vs. remote sampling?
18
18 Applications Sharing cost of assertions Isolating deterministic bugs Isolating non – deterministic bugs
19
19 Experiments Sharing the cost of assertions Sample assert ( ) statements in CCURED code Overhead incurred: ScenarioAverageMaximum Unconditional55%181% 1/100 sampling17%46% 1/1000 sampling10%26%
20
20 Effectiveness of Sampling At density 1/1000 for observing rare program behavior? To achieve confidence level =90%, Least number of runs needed = 230,258 ! Solution: No. of licensed Office XP users = 16 million 2 runs/week/user = 230258 runs every 19 min!
21
21 References Sampling User Executions for Bug Isolation Public Deployment of Cooperative Bug Isolation Scalable Statistical Bug Isolation www.cs.wisc.edu/liblit
22
22 Thank You
Similar presentations
© 2025 SlidePlayer.com Inc.
All rights reserved.