Do we need to and how to simulate real user's pattern for the performance load?

I saw lots of people are talking about simulating the real user load according to the real production site for better analyzing or handling the performance status/issues they have. This is re开发者_JAVA技巧ally cool because we can easily found the performance problems and solve them in the "dawn of the day" when new features or updates we've done.

And here is basic workflow for the round trip:

1. Get the user's real load in access log and find the corresponding detail request from production
2. Store these real load information and map them into some fast storage (my first thought is mongodb or something like that)
3. Start up multi-threads to push the recorded loads into local performance env.
4. Lets see what will happen then.....

So my first question is what are you thinking of the "simulate real user's pattern", are they really valuable or just wasting time?

I know Jmeter has a AccessLogSampler for simulating the http request but mine is soapRequest to the webservice. Looks I have to do lots of implements in java but any other chocie?

Your ideas are really really appreciated!

Vance

---

If you don't simulate real load, then you won't find real performance problems.

True story: I once developed a new feature in a database-centric application. I wrote performance tests based on how I thought the feature would be used, and optimized accordingly. Then it was released, and used in ways that I didn't imagine -- and most of my optimizations were worthless. I would have been better off simply implementing the un-optimized feature, and waiting for real data.

---

Web services tests have 2 major types: hit-based and scenario-based. Hit-based is very simple and concentrate on distributions/request rates. Scenario-based concentrate on sequence and logic / transaction rate / concurrent virtual users.

Hit-based is much simplier to load test and this way is preferred for really high-load systems. Scenarios unavoidable sequence-dependent in web applications.

Everything depends on your service (or part of service).

My alternative to AccessLogSampler is HTTP Raw Request