PI: Katerina Goseva

1
Performability of Web-based Applications

• PI Katerina Goseva Popstojanova
• Students Ajay Deep Singh Sunil Mazimdar
• Lane Dept. Computer Science and Electrical
  Engineering
• West Virginia University, Morgantown, WV
• katerina_at_csee.wvu.edu

2
Problem

• World Wide Web is the biggest existing
  distributed system so far
• Huge number of Web clients - tens of millions and
  rising
• Users demand 24/7 availability and response time
  within several seconds
• However, very often they experience long and
  unpredictable delays

• (WWW - World Wide Wait)

• Problem Traditional analysis and prediction
  methods do not work for Web

3
Relevance to NASA

• Increasing use of Web-based technology at NASA
• Web sites Agency wide
• Control of daily mission operations from multiple
  geographically distributed locations via Internet
  (e.g., Web Interface for Telescience at JPL)
• Real-time applications remotely
  controlled/monitored over the Internet or an
  Intranet (e.g., Tempest embedded Web server at
  Glenn Research Center)

4
Relevance to NASA

• Our empirical analysis is based on data extracted
  from actual Web logs of ten servers
• Three public and three private Web servers at the
  NASA IVV Facility

• Lane Department of Computer Science and
  Electrical Engineering (CSEE) Web server

• NASA Kennedy Space Center (NASA-KSC) Web server

• Campus wide Web server at the University of
  Saskatchewan
• Web server of the commercial Internet provider
  ClarkNet

5
Approach

• Develop methods and tools that are general and
  powerful enough to provide flexible analysis and
  quality assurance of Web reliability,
  availability, and performance
• Develop scalable framework that combines
  measurements and models at different levels of
  detail and abstraction
• Reliability/Availability based on typical usage
  patterns
• Performance non-Poisson queuing theory
• Combine reliability / availability and
  performance and analyze their tradeoffs

6
Approach
User session characterization
Web access log analysis
Realistic workload
Session layer (user view)
Performance model
Software/hardware resource utilization
Service layer (software architectural view)
Performability model
Application hardware resource monitoring
System layer (deployment view)
Software/hardware failure/recovery characterizatio
n
Reliability/ availability model
Resource layer (hardware device view)
Web error log analysis
Request-based and session-based error
characterization
7
Accomplishments
Create relational database
8
Accomplishments

• Empirical analysis of the Web workload, errors,
  request-based and session-based reliability for
  ten Web servers

• Some examples

• Fixing the errors with the highest frequency of
  occurrence is the most cost effective way to
  improve Web quality

9
Accomplishments

• We argue that session-based reliability is a
  better indicator of the users perception of the
  Web quality than request-based reliability

10
Importance/benefits

• Innovative theoretical and empirical research
  results
• Introduced and empirically analyzed new measures
  for session-based workload and reliability
• Conducted detailed empirical study on Web errors,
  including severity level of errors, unique
  errors, and unique files with errors measures
  that have not been considered earlier

• Practical value
• The results of our research were actually used by
  Web administrators of the NASA IVV and CSEE Web
  servers to improve their quality in a
  cost-effective way

11
Next steps

• Performance attributes
• Develop non-Poisson queuing theory
• Standard performance models (Queuing Networks
  Layered Queuing Networks) assume Poisson arrivals
• Web workload is bursty (highly non-Poisson)
• Dependability attributes (reliability,
  availability)
• Develop architecture-based models based on
  typical usage patterns
• Combine performance and reliability /
  availability models
• Analyze tradeoffs among multiple quality
  attributes