For large result sets, (5000k+ TestDocuments) I'm seeing deserialization of MLPP slower than for the default provider. Hooking up the profiler, the issue appears to be just getting the results of the Aggregation Framework command. Getting those results is using the BsonBaseSerializer whereas the default linq provider is using the BsonClassMapSerializer. Perhaps that's the issue? See the attached screen shot of a profile analysis.
(Moved this issue over from bitbucket)