1) Stress-testing scenarios should model user’s behavior as closely as possible (hence all that xml configuration in stress-testing tool)
2) Performance whack-a-mole game is a non-ending one, you’d better put yourself some deadlines where to stop at
3) Real-model performance testing is way more accurate than general capacity planning guidelines. So if you’ve got heavy hardware at stake: consider doing a preliminary load-testing, it can reduce potential hardware cost. It was multiple times in our case )
4) TM1 9.5.2 Parallel Interaction feature really works, especially in models with lots of rule calculations. We’ve got 1,5 performance gain on 100 concurrent writing users just by turning it on. And yes it takes 20-30% more memory, sigh.
Overall, watching the CPUs throttle up on big AIX machine makes my internal geek happy.
It was a very interesting project to do, I don’t think a lot of people emulated a 1000 users load on TM1 model. Although, I personally consider such user load unreal for a planning system, but if the customer says so — we test it.