E-Business Capacity Planning

E-Business Capacity Planning

Capacity planning forecasts when future workloads may saturate an e-business system, then identifies cost-effective ways to delay saturation while maintaining acceptable service.

Key framework and process

• Forecast future workload based on:
  • Natural growth in existing traffic as the site becomes better known.
  • New applications or services that attract additional customers.
  • Changes in customer behavior, including seasonal peaks and promotions.
• Set service-level requirements. Management should define acceptable response times, throughput, and availability based on the business and its customers.
• Assess service quality and failure risk. Estimate the cost of downtime, identify single points of failure, and consider reliability measures such as mean time to failure and mean time to repair.
• Model workload and identify bottlenecks. Trace customer actions through the systems they use, estimate resource demand, and find the functions or components that constrain performance.
• Plan how to scale. Scale up by increasing a server’s capacity, or scale out by adding servers to a cluster.
• Compare capacity costs with the cost of service degradation. Decide whether to invest in permanent infrastructure or use temporary capacity for short-lived peaks.

Performance modeling can use queueing models. A customer operation, such as product search, can be treated as a queue. A more detailed model can represent the web, application, and database servers as a queueing network. These models can estimate utilization, throughput, waiting time, and response time.

Service levels depend on the business

• An online retailer may lose customers to competitors if its site is slow or unavailable.
• A digital library hosting unique research papers may face less immediate customer switching because the content is not available elsewhere.
• The lecture cites the “eight-second rule” as a general warning that users may leave if a response takes too long. It gives less than two seconds of server-side response time as an example target, while noting that end-to-end response time also depends on the network and client device.

Example: Planning for a sales peak

The lecture uses access-log data for a search operation:

Measure Normal workload Fourfold peak Completed search requests 18,000 in 3,600 seconds — Arrival rate 5 requests/second 20 requests/second Average search execution time 48 ms — Server utilization 24% 96% Average response time 0.063 seconds 1.2 seconds

At normal traffic, the modeled server has spare capacity. At four times the arrival rate, utilization rises to 96% and response time increases sharply. Average utilization can hide bursts in arrivals, so a system may fail under variable peak traffic before reaching 100% average utilization.

Business decision: For a brief annual sale, temporarily renting or otherwise adding external server capacity may be more cost-effective than permanently expanding infrastructure. If peaks occur frequently, investing in additional capacity may make more sense.

Actionable recommendations

• Use access logs to estimate request arrival rates and service times for important customer actions.
• Forecast seasonal and promotional peaks rather than planning only for average traffic.
• Set performance and availability targets that reflect customer expectations and how easily customers can find substitutes.
• Test whether modeled peaks push utilization or response time beyond acceptable limits.
• Identify single points of failure and bottleneck resources before scaling.
• Compare permanent infrastructure investment with temporary capacity for occasional events.

Presenter and source

• Presenter: Unnamed lecturer
• Source: E-Business Capacity Planning, E-business channel