Well, this topic could fill an entire book...
If you implement a model the first question you should as is: What is the purpose of the model? Which questions should the simulation of the model answer?
Looking at architectural exploration which goes quite often hand in hand with performance analysis the question is: does my HW/SW split and my HW partitioning satisfy my perfomance requirements (wrt. latency, thru-put, compute.efficiency, power,...). In this case you usually do not need to implement a particular functionality in detail rather something that 'behaves like' in terms of your requirements. E.g. if you need to check that your communication scheme (buses, arbiters, bridges etc.) fulfills the needed band with you use traffic generators but have a fairly accurate bus model, sometimes even at AT. And you need to implement the mechanisms to observer and extract the needed performance indicators to allow the analysis
For software development the requirements are different. Here the maximum simulation speed is required so whereever possible you take short cuts. Bus transaction are not modelled anymore rather DMI is used (of course if functionality allows to do so e.g. when reading from/writing to a memory) and the entire model may run in LT mode which allowes parts to independently advance in time. Peripheral units may be modelled register-accurate but with out real functionality, i.e. a system control unit does not follow the needed scheme if changing e PLL frequency and alike.
This might give you some high-level clue. There are many more things to it but all of them depend on the answer to the initial questions.
Maybe the DCVon Europe 2017 tutorial on virtiual protorypes might provide a few more answers. You may find a PDF version of it at the MINRES site in the Publications and Papers section or at https://minres.com/downloads/VP_Tutorial_DVCon-2017.pdf as well as at the DVCon Europe website https://dvcon-europe.org/conference/history