Clustering Windows Applications based on failure behavior

1
Clustering Windows Applications based on failure
behavior

• Archana Ganapathi
• (archanag_at_cs.berkeley.edu)

2
What Data do I have?
3
Crash Cause (app)
4
Project Goal

• What is the best way to partition/cluster
  applications based on their failure behavior?
• Examples
• Type of application
• (e.g. web browsing, text editing, e-mail, )
• Who wrote the app
• (e.g. Microsoft, Mozilla, )
• Shared components (e.g. MS office products)

5
Sample Clusters by app type
6
Sample Clusters by developer
Microsoft
Mozilla
IEMS WordOutlook
Firefox Thunderbird
Mathworks
MatlabSimulink
7
Process (1)

• Data Representation what metrics matter?
• per-app vector
• elements per-machine stats?
• A1 avg. time between events in a crash, A2
  avg. time since last crash
• Distance Measure to compare app failure behavior
  similarity
• L1/L2/Linf norm?

8
Metric (time between crashes)
9
Metric (time between crashes)
10
Metric (usage vs crashes)
11
Metric (usage vs crashes)
12
Web Browsing
13
email
14
Process (2)

• Application Clustering rules
• What is the distance threshold for
  distinguishing related and unrelated application?
• ab, bc gt(?) ac
• Evaluation
• Compare to hypothesized clusters
• Hidden-relationships between apps

15
Document presentation/prep
16
Code development
17
multimedia