And to futher beat dead horses, lets make sure you understand this little esoteric bit of knowledge. One way to add 9's to the availability of service x is to have more than one server acting as if it were one box. Note that this generally means that you'll need additional hardware to pull off the magic that makes this occur, but it is also possible just using open source clustering software.

For this analysis, you need to know what your current MTBF and MTTR for the server and software responsible for service x. Almost nobody actually collects that data before heading down the path you're on now, so you'll have to guess. Assume you guess that the Mean Time Between Failures and the Mean Time To Repair numbers end up giving you an availability of 94% per year. If we further assume that you are able to add an exact replica of that server/service and ignore the additional failure points added by the clustering software/hardware, then to find the new availability just multiply (1-0.94) by (1-0.94) and you get .0036, subtract that from 1 and you get 99.64% availability.

As you split things up, it may get easier to manage, and easier to replicate the various pieces, but all this means you get to complicate the picture more and more, and as the server/service availability gets better, the more you need to ensure the networking parts of all this are equally reliable/available/fault tolerant. Then the supply of power get to be important, etc.

-Scott