Load Prediction & Code Profiling for HR Workforce Portal

Situation

Harbinger's customer provides HR communication and benefits content to organizations through a web-based portal. The product is used before the annual employee benefits enrollment period and usually requires extensive research on the part of employees during the run up to the enrollment period.

Challenge

The company had started signing up larger-sized businesses consistently, and had recently closed a deal with an industry leader with about 80,000 employees in the US. The larger companies also wanted a shorter enrollment period of 2 to 3 weeks in place of the usual 2 months, resulting in higher user footprint per hour on the portal. The enrollment period was just a quarter away, and the company sensed an urgent need to get the portal ready for the large number of employees from these big-ticket accounts.

Harbinger Solution

Harbinger carried out an analytical exercise to predict load using Microsoft SQL Server 2005 Analysis Service. This tool provided a parametric model based on three years of historical usage of the system. The output from the SQL Server Analysis Service was used to build a statistical spreadsheet model using prediction theories like time series and decision tree. The spreadsheet model predicted future usage trend and also the hourly and daily load on the system on each subsequent day of the enrollment period. The analysis indicated that the system would need to support up to 3000 concurrent users per hour.

In the meanwhile, another team of engineers was preparing test scripts and test data in Microsoft Visual Studio Team System 2008 (VSTS) to simulate load scenarios and to benchmark system performance. Harbinger profiled the server code and SQL scripts to identify bottlenecks in the system and determine optimization opportunities, which were incorporated in tandem.

• Microsoft SQL Server 2005 Analysis Service
• Microsoft Visual Studio Team System 2008 (VSTS)
• Statistical spreadsheet modeling using time series and decision tree prediction theories

Benefits

The system went beyond the targeted numbers on every count. The production environment eventually had a web farm of eight load-balanced servers that served about 9000 users per hour without stressing the system. The SQL parametric model and the spreadsheet model captured the hourly and daily usage trend to predict the peak load hour and day of the enrollment period very accurately. Despite giving only the minimum three years of historical data to it, the model predicted usage within an error margin of 10%. With one more year of usage statistics, reusable models and all code profiling work behind it, the company continues to add customers and proactively plan for load in the coming years.