The answer to that is complicated. In order to avoid more complicated answers most people recommend tuning every node of the network to the same MTU value (where any MTU > 1500/1514 is generally considered to be jumbo frames).
In reality, if you have a subset of your network nodes communicate using jumo frames and the rest communicate using "standard" MTUs then it will work just fine. You just have to ensure that the switches/routers that make up the transport part of your network support the maximum MTU you are using.
The problems will arise when the machines using jumbo frames try to interact with the machines that do not support jumbo frames. In some cases (most cases, actually) everything will be fine the layer-4+ protocols will just negotiate to the maximum common MTU value. Some things might take a bit longer because of this extra negotiation and some protocols may never reach maximum performance due to the mismatch (for example, TCP congestion controls will take longer to converge onto the best window sizes). But most of the time you wont notice. A few things might not work at all (e.g., UDP based streaming from the "larger" MTU server).
Of course, you also need to examine why you are looking to do jumbo frames at all. In GigE environments using modern NICs (i.e., NIC with protocol offloads) there is little to no advantage to jumbo frames. If you are doing 10Gbe then you might see some gain (though even that is subject to some fair debate).