January 17, 2000
|
Printer ready |
At Living.com Inc., an E-commerce site that sells furniture and other household goods, senior performance analyst David Langworthy sees no advantage in using artificial workloads created by manually clicking through the company's development site. "The question the execs want me to answer is: Can we handle a certain capacity? These commercial load-testing tools do not help answer that question," Langworthy says.
Instead, he created a set of Perl and Csh scripts for stressing the company's Web application by using actual workloads obtained from Web-server logs.
"What you need to do for capacity planning is know how many real users are on the system," Langworthy says. "It's next to impossible to get a correlation between a synthetic load and a real load."
Using the Web-server traffic logs from Living. com's front-end Netscape Web server, Langworthy's scripts, used in conjunction with a Java playback tool provided by Art Technology Group Inc., extract a list of user events that happen at peak times of the day and play them back against Living.com's Web application. The result is a real workload stress test reflecting actual user traffic.
Although the Netscape Web-server logs provide adequate user-event information, more detailed user-state information would provide a more accurate stress test. In order to get detailed user traces from Living.com's Netscape Web server and Art Technology Group Dynamo application server, senior Web developer Holland Gary created a logging system that saves all the users' state information necessary to recreate the user actions accurately on the system. Although the Art Technology Group Dynamo application server has logging capabilities, Gary wanted to create a unified logging system with specific capabilities.
"It's a matter of control," Gary says. "We wanted to make sure the logging system didn't impact performance and had failover capability." Gary's logging system is also portable, so that it can be moved to another application server, if necessary, with only minimal changes. It allows better identification of users as well. With many users on the Internet coming from sites that use Network Address Translation, such as America Online, there are often different users using the same IP addresses.
Also, Living.com uses stored cookies so the logged URLs don't always reflect unique users. Gary's logging system has the additional information necessary to distinguish these individual users correctly. Future versions of Langworthy's stress-testing tools will make use of these enhanced logs to provide an even more realistic workload.
oftware that generates artificial workloads using virtual users is fine for testing the scalability of sites that aren't online yet. But if your site is already online, then you may already have real users creating real workloads. The "click" logs or other audit trails generated by your Web and application servers are a transcript of real, not artificial, workloads. Because they reflect what real users do with your site, it's clearly the best data with which to test your site's scalability.
Return to main story, "Is Your Web Site Scalable Enough?"
Back to This Week's Issue
Send Us Your Feedback
Top of the Page