In most tests, changes to this parameter make little difference to end-to-end performance.
Things to check:
1. Do you really need to use Merge Data Streams? You should only ever use this if you are merging from multiple readers - NOT from multiple paths from the same reader (where any standard processor can unite paths). Unfortunately, it is very often misused.
2. Could you re-configure your process such that the Group and Merge acts only on the columns it really needs to and you use a Lookup and Return for the rest? This would avoid passing so many columns through and would likely radically improve performance.