Alright, so there isn't a clear culprit, but there's a good chance it's the heuristics for STUN - after 1.8 a conversation lookup was added (r44068), and that uses a g_hash_table. The only major difference between the two profiles that I found was a significant increase in calls to g_hash_table_lookup from the conversation code.
If you don't need STUN decoding, try disabling its heuristic (there's no preference, but just commenting out the heur_dissector_add call should work) and see if that improves things.
If that doesn't work I'm not sure what the problem could be.
Evan
PS the new name-resolution code also uses g_hash_tables now, but the call source is comparatively small next to the conversation code, so I don't think that's the problem. It may be worth trying a 1.10 build as well though, just for comparison (since that will have the STUN changes but not the name-resolution changes).