Dynamic SemiMarkovian Workload Modeling

1
Dynamic Semi-Markovian Workload Modeling
Nima Sharifimehr, Samira Sadaoui Department of
Computer Science, University of Regina, Regina,
SK Canada S4S 0A2 sharifin, sadaouis_at_cs.uregina.
ca
2
Outline
1. Motivations 2. Solutions 3. Dynamic
Semi-Markov Model (DSMM) 4. Dynamic Building
Process of a DSMM 5. Integration into Enterprise
Application Servers 6. Evaluation Results 7.
Summary 8. Future Research
3
1. Motivations

• Performance of SS affected by incoming workload
• Performance of SS evaluated through workload
  analysis
• Workload analysis through workload modeling
• Markovian approaches to model workload for a SS

4
1. Motivations

• Performance analysis of Enterprise Application
  Servers (EAS) is critical for all e-business
  enterprises
• Lack of workload modeling for EAS

5
2. Solutions

• A dynamic semi-markov model (DSMM) for an
  accurate workload modeling
• DSMM to efficiently model workload of enterprise
  application servers

6
3. Dynamic Semi-Markov Model
A combination of Semi-Markov Model and Dynamic
Markov Model

• P probability of evolving the system from one
  state to another when an element is seen in the
  analyzing data
• PDFt evolution speed from one state to another a
• PDFs idle time spent when a transition evolves
  the system

7
An example of DSMM
8
4. Dynamic Building Process of a DSMM

• A DSMM changes dynamically to reach the best
  model for a system
• Modifying the dynamic modeling approach used in
  Dynamic Markov Compression (DMC) to be applicable
  for the semi-markov model

9
5. Integration into Enterprise Application Servers
- RUBiS module generates workloads with high
randomness which simulates chaotic situation -
GMC builds a DSMM-based workload model for an EAS
10
6. Evaluation Results
Measuring the accuracy of DSMM for EAS
Similarity of built DSMMs and real workloads
11
7. Summary

• Formal definition of a Dynamic Semi-Markov Model
  (DSMM)
• An algorithm for building a DSMM for any
  application
• DSMM as an efficient workload modeling tool for
  enterprise application servers

12
8. Future Research

• Defining an operator that measures the similarity
  of two DSMMs to reduce the repository of DSMMs
• Investigating the applicability of DSMM on other
  applications