1. 1 Challenges in Scaling E-Business Sites  Menascé and Almeida. All Rights Reserved. Daniel A. Menascé Department of Computer Science George Mason University

2. 2 Impacts of Bad Performance  Bad performance: response time above 8 seconds (eight-second rule).  $43.5 billion lost each year in e-commerce due to bad performance (Zona Research, April 1999).  Holiday Season of 1998: over 1/3 of customers gave up due to slowness, 44% turned to conventional stores, 14% moved to another site.  Menascé and Almeida. All Rights Reserved.

3. 3 Performance Problems for E-commerce tend to get worse!  Proliferation of mobile devices  Easier to use interfaces (VUI, wireless and Web services on cars and airplanes, novel browsing paradigms)  Increasing load placed by agents  Impacts of authentication protocols (e.g., TLS) on e-commerce site performance.  Menascé and Almeida. All Rights Reserved.

4. 4 Typical Questions  Is the online trading site prepared to accommodate a 75% increase of trades/day?  Do I have enough servers to handle a peak demand 10x the average?  How fast can the site architecture be scaled up? What components should be upgraded? Database servers? Web servers? Application servers? Bandwidth?  How can I design a site that will meet its business goals?  Menascé and Almeida. All Rights Reserved.

5. 5 Outline  Scalability  A Reference Model for E-Business  Workload Characterization  Customer behavior model graphs  Client/Server Interaction Diagrams  Concluding Remarks  Menascé and Almeida. All Rights Reserved.

6. 6 A Reference Model for Electronic Business Business Model Functional Model Customer Model Resource Model Characteristics of the Business Navigational Structure of the Site Patterns of Customer Behavior Site Architecture and Service Demands External Metrics Internal Metrics Business View Technological View Descriptors  Menascé and Almeida. All Rights Reserved.

7. 7 External Metrics and Descriptors  Cover the Nature of Business  Metrics:  Revenue throughput (dollars/sec)  Potential lost revenue/sec  Click-to-look ratio  Look-to-basket ratio  Basket-to-buy ratio  Click-to-buy ratio  Availability  Download times  Page views/day  Unique visitors/day  Menascé and Almeida. All Rights Reserved.

8. 8 External Metrics and Descriptors  Descriptors:  Number of registered customers  Number of potential customers  Maximum number of simultaneous customers in the store  Number of items in the catalog  Estimated operational cost  Menascé and Almeida. All Rights Reserved.

9. 9 Workload Characterization 1.Determine the e-business functions made available by the site. Associate URLs or URL patterns to each e-business function. 2.Analyze the site’s HTTP logs to determine customer sessions. 3.Cluster customer sessions into groups of “similar” customer sessions, i.e., sessions that exhibit similar behavior.  Menascé and Almeida. All Rights Reserved.

10. 10 home browse search addselect pay 0.5 0.3 0.35 0.15 0.2 0.3 0.1 0.2 0.1 0.4 0.1 1.0 entry 1.0 Customer Behavior Model Graph  Menascé and Almeida. All Rights Reserved.

11. 11 Metrics Derived from the CBMG  Average Number of Visits Per State  E.g., average number of searches per visit to the site,  Average Buy (or open account) to Visit Ratio – also called conversion ratio.  Average Session Length Per Visit  Menascé and Almeida. All Rights Reserved.

12. 12 Metrics Derived from the CBMG  Menascé and Almeida. All Rights Reserved.

13. 13 Workload Characterization Methodology Merge and Filter Get Sessions Get CBMGs (clustering algorithm) HTTP Logs Session Log CBMGs Request Log  Menascé and Almeida. All Rights Reserved.

14. 14 Result of Clustering Analysis Cluster 1: majority of sessions, short sessions, and highest BV ratio. Cluster 6: small fraction of sessions, large sessions, smallest BV ratio.  Menascé and Almeida. All Rights Reserved.

15. 15 Buy to Visit Ratio vs. Session Length  Menascé and Almeida. All Rights Reserved.

16. 16 Architecture of E-Commerce Sites  Menascé and Almeida. All Rights Reserved.

17. 17 Remarks  For e-commerce, need to characterize the workload at higher levels of abstraction: sessions vs. requests.  Workload characterization has to be customer behavior-oriented as opposed to request-oriented.  Customer Behavior Model Graphs capture customer behavior and can be mapped to resource demands.  Menascé and Almeida. All Rights Reserved.

18. 18  Capacity Planning and Performance Management:  key to EC: the competitors are just a click away!  require predictive models: avoid ROTs!  throwing more plumbing is not the solution.  models have to integrate the business, customer, and resource aspects of the problem. Remarks (cont’d)  Menascé and Almeida. All Rights Reserved.

19. 19 Performance Models of E-Commerce Sites Daniel A. Menascé Dept. of Computer Science George Mason University

20. 20 Component-level models server incoming link outgoing link server

21. 21 Component-level models server cpu disk 1 disk 2 incoming link outgoing link server

22. 22 Component-level models server cpu disk 1 disk 2 incoming link outgoing link server

23. 23 Component-level Models  Each component is represented by a resource (e.g. CPU, disk, communication link) and a queue of requests waiting for the resource. resource queue

24. 24 Basic Concepts  Utilization of a resource:  Fraction of time the resource is busy serving requests during a measurement interval.  Example: the CPU was busy during 40 minutes during a measurement period of one hour. It’s utilization is then:  40/60 = 0.667 = 66.7%  Utilizations are measured as dimensionless numbers

25. 25 Basic Concepts (cont’d)  A request may be using a resource (e.g., CPU, disk, etc) or waiting to use it.  The time spent using the resource does not depend on the number of resources waiting to use the resource.  The time spent waiting to use the resource depends on the load, i.e., on the number of requests in the queue waiting to use the resource.

26. 26 Performance Model Parameters  Workload Intensity  HTPP Requests/sec  Transactions/sec  E-business functions/sec  Service demands for each resource and each type of request.

27. 27 Service Demands cpu disk 1 disk 2 incoming link outgoing link server 0.109 sec 0.00107 sec 0.003 sec 0.08 sec 0.12 sec Service demands do not include any queuing time! It is just service time.

28. 28 Computing Waiting Times cpu disk 1 disk 2 incoming link outgoing link server 0.109 sec 0.00107 sec 0.003 sec 0.08 sec 0.12 sec Waiting times depend on the load (arrival rate of requests) and on the service demands.

29. 29 DiDi RiRi Service demand at resource i Utilization of resource i (U i ) Computing Residence Times Residence time at resource i

30. 30 Residence Time at Incoming Link cpu disk 1 disk 2 incoming link outgoing link Web server 0.109 sec 0.00107 sec 0.003 sec 0.08 sec 0.12 sec  req/sec

31. 31 Residence Time at Outgoing Link cpu disk 1 disk 2 incoming link outgoing link Web server 0.109 sec 0.00107 sec 0.003 sec 0.08 sec 0.12 sec  req/sec

32. 32 Residence Time at the CPU cpu disk 1 disk 2 incoming link outgoing link Web server 0.109 sec 0.00107 sec 0.003 sec 0.08 sec 0.12 sec  req/sec

33. 33 Residence Time at Disk 1 cpu disk 1 disk 2 incoming link outgoing link Web server 0.109 sec 0.00107 sec 0.003 sec 0.08 sec 0.12 sec  req/sec

34. 34 Residence Time at Disk 2 cpu disk 1 disk 2 incoming link outgoing link Web server 0.109 sec 0.00107 sec 0.003 sec 0.08 sec 0.12 sec  req/sec

35. 35 Summary of Results Average Response TimeSum of service demands

36. 36 Response vs. Arrival Rate

37. 37 Example: online trading site  Open QN for the online trading site: