Hi Max,
Theoretically, for a small number of elements, usually somewhere between 3 and 30 depending on implementations, a linear search is faster than a hash search, especially in the Pharo dictionary hash search implementation.
Efficient dictionary implementations are usually bucket-based. The dictionary holds a certain number of buckets, and based on the key hash, the bucket where the key value is present is determined. Small buckets are linear (arrays or linked list). Large buckets are typically balanced binary trees (red-black trees typically). Under a certain number of elements there is a single bucket, which means a linear search is performed, as for the SmallDictionary. When it grows the dictionary search becomes a combination between a hash search and a linear or tree search.
Pharo dictionary search is first hash-based, then all the buckets are represented next to each other in the same arrays and a linear search is performed there, leading to many collisions and slower search time (especially when the element is not found), sometimes the code searches over multiple buckets because the dictionary is too full or there are too many near-collisions. The memory consumption is competitive with the advanced implementations though (precise measurements would need to be made).
Method dictionaries are represented differently to optimize the look-up logic.
If you want to improve things and have one dictionary implementation instead of two, implement or look for a bucket based dictionary and put it in the base image instead of Dictionary. This is quite some work since there are many APIs to port. You can look at the Pinnochio implementation, it's quite good but they've not implemented large buckets.