I've no clue about MongoDB performances, and which of Redis and MongoDB is faster incrementing integers, but for sure this benchmark is completely useless, as it is actually measuring a mix of client library latency and round trip time, being it single threaded, making no assumptions about the durability requirements, and so forth.

For sure it seems like that MongoDB client library you are using has a smaller latency, perhaps it is a C library? As Redis-rb is instead written in Ruby itself.

Another huge problem of this benchmark is that it should try to increment a random key per every iteration, out of a dataset with a few millions of keys, as this is realistically what you need. You'll see how Redis with 50 clients will increment this counters 100k times per second or more without any performance degradation as the number of keys grows.