I'm a firm believer in working in a very small dataset while exploring, experimenting, and testing alternatives. Having a small sandbox for doing an experiment cuts the turnaround and simplifies controlling variables. After building a nice DTS package to create a sandbox database, I just realized how easy it is to do it directly in AS. For two reasons.
First, AS has a new layer called a Data Source View (DSV) that basically buffers the logical metadata. I just added WHERE (ProductKey % 50 = 1) to the DSV table definition for the Product dimension and instantly got a small subset. Of course, you could always do this by creating a view--but that would then require making changes within the source database and also switching the table source in the AS designer. By doing it in the DSV, it's isolated from the SQL database and doesn't require any changes to anything.
Second, AS has a new “Unknown Member” concept. In 2000, discarding dimension members would have either discarded all the orphaned fact rows, or just failed with “dimension member not found”, depending on how you built it. In 2005, you can lump all the fact rows with invalid dimension members into an Unknown bucket for each dimension. This means a) that the totals tie out to the original data source, and b) subsetting multiple dimensions still leaves useful combinations--often with the Unknown members of other dimensions. Even a relatively large fact table processes very quickly with small dimensions, so this doesn't mess up the “quick turnaround“ benefit of the sandbox, and if it did, you could just use the DSV to subset the fact table too.
It's also really easy to keep two copies of the DSV - one filtered, one not - so I can switch back to the full dataset version with no problem.