What methods do you use to test for scalability in web applications?

Question

Our testing system is pretty rudimentary; fire up a browser, see if it works. Recently we ran into problems, found by our client, with our application where the number of users created a slow-down in the application. The application is basically a huge Word document with people editing their own versions all at the same time. Part of the problem came from not knowing how to test multiple instances at the same time. My partner and I thought about how to test this; one idea was to hire out an internet cafe and hire students for an hour to bang on the app.

What are other ways that people have tried to emulate concurrency in testing their web-based application? Most of the advice here is for specific methodology; I'm asking, how do you test it to make sure that it works?

Answer 1

If you have never checked out Selenium, then you need to. It will allow you to do automated web testing through the browser. Ok, so first problem solved.

Now ideally you could use that same script and load it up on a bunch of boxes and run them all at once to get some sort of load testing right? Luckily for you someone has already figured this out, although it is a paid service: Browser Mob. But, it looks like you were willing to spend a little money to do this anyway, and would probably net you better, more repeatable results.

Answer 2

JMeter http://jakarta.apache.org/jmeter/

Answer 3

We usually answer the question "can the web application do more than one thing at a time" by using JMeter to produce a simulated HTTP load on the web server.

I find that it helps to consider distinguish several different types of testing; concurrency (what happens when two events in the system collide), capacity (what happens when there are many overlapping requests), volume (what happens as data accumulates in the system)...

Huge general slow down, evidenced by response times that fall outside of the SLA, are usually related to capacity problems (with contention as a common cause) or volume (many users, much data, and the system gets slower over time). The former usually requires some sort of multi-threaded request stream; the latter you can usually manage by preloading the volume, and then measuring the response times experienced by a single user.

I generally find that separating the load generator from the actual measurement/instrumentation is a good idea. That can be as simple as having a black box over there to generate a typical load, and sitting here with a stop watch measuring the responsiveness of a typical use case.