Draw up a diagram showing the components of the system indicating the two most important metrics across each connection and within each component.
Then, for each metric establish how much they affect each other, and are affected by the parameters you are looking at. Determine factors by which each of the components are affected by the required metrics and the various connected measures.
This will give you a first cut evaluation of the requirements.
Then, build a test environment and use this to prototype the system and improve the calibration of the relationships in the parameters with sub-scale experiments. This is also essential for routine stress testing before you need to commit to a final configuration.
Once deployed, continue measuring to see that your system continues to perform as designed. You can often see a variance in the coarse overall metrics way before any other detailed parameter becomes visible.
Short answer - Unless you have a fully scaleable application, you will need to guess and add contingency. Scaling is seldom truly linear, and the effects of multiple parameters can be really counter-intuitive. Trying to do this without a calibration model is extremely hopeful.
