Regulating Data Flow in J2EE Application Server

1
Regulating Data Flow in J2EE Application Server

• Wei Xu (xuw_at_cs.berkeley.edu)
• Zhangxi Tan (xtan_at_cs.berkeley.edu)
• Armando Fox
• David Patterson

2
Motivation

• Regulate workload flow in software systems
• Workload pushed through multiple tiers
• Black-box components
• Different processing time and resource
  requirement in each tier
• Unpredictable disturbance
• Sometimes a safety problem instead of performance
  problem
• Examples
• Streaming database (last retreat)
• J2EE application

3
Problem and Control Loop

• Database overload gt deadlock
• Need to regulate incoming workload at the web
  server
• Prevent client visible error

4
Problem with simple control

• Each HTTP request causes different load on db
• Too noisy for controller
• gt Need better request -gt db prediction

5
Classifying workload

• Classifying workload at web server
• k-means
• Small/large request
• by approx. db usage
• Offline profiling, but can be online

6
Dynamic Probabilistic Scheduling

• Naive admission control punish small requests
• Underutilize web server tier
• Give small requests higher priority
• Keep fairness to large requests
• Need to adapt to changes in workload

7
Conclusion

• Using control theory to regulate dataflow
• Black box components (different from traditional
  software system flow control)
• Components cannot drop data (different from
  networking)
• Algorithms can be used with little modification
• With a little application specific knowledge, can
  avoid complex model
• Most of the parameters tuned automatically
• Future work
• Does workload classification work in more complex
  applications
• Online classification without profiling
• Interact with resource allocation