UPDATE 2012-03-16: Please also take a look at Slowly Changing Dimensions with MD5 Hashes in SSIS which we have determined to be fastest, most efficient approach to maintaining Type 1 dimensions.
A few days ago, one of our SSIS packages that maintained a Type 1 Slowly Changing Dimension (SCD) of about 1 million rows crept up to 15 minutes of runtime. Now this doesn't sound too bad, but this is part of our hourly batches, so 15 minutes is 25% of our entire processing window. The package was using the Slowly Changing Dimension Wizard transformation - we were doing the standard OLEDB Source (which basically represented how the SCD "should" look) and then sending it to the SCD transform and letting it figure out what needed to be inserted and updated. One option was to switch to lookups instead of the SCD wizard to speed things up, maybe even some fancy checksum voodoo for the updates (see http://blog.stevienova.com/2008/11/22/ssis-slowly-changing-dimensions-with-checksum/ for an example). Then after thinking about it a little more - why are we sending a million rows down the pipeline every hour? We know only a small percentage of these are new - and another small percentage needs to be updated. Well, we can just write a quick SQL query to get us just those sets and the package would be much more efficient!
Wait a tick - why would we give the rows to SSIS if all it is going to do insert one set and update the other? Let's just do it all in T-SQL: