Look at this from a pure test design issue. In a test you want to minimize the uncontrolled factors in your test because this impacts both the in-test and initial conditions for the test and also impacts repeatability of the test as an output. On a process basis I recommend a minimal network for the majority of the load and perhaps a control element (single virtual user) on a remote basis to check the performance of the network. If the user has network performance testing tools available, then actually I go for parallel tests, one on the application and the other on the network with the data flows from the application to measure pure network influence on response times.
Once I understand the overall scalability of application then I bring in the last remaining uncontrolled element, network, since the application at that point is a "known good" at all load levels. Then I look at the influence of network under load. This assumes that I do not have any of the WAN emulation options available to me from Shunra to do the modeling within my own lab (which would be preferred).