I would have your response time messages sent to a topic and every node which needs this information take a copy. When a node need to know which service to use, it looks at its local copy. This way the lookup is very fast and the nodes can run independently.
BTW You may need to periodically check the latency of the slower systems. i.e. once a system appears slow it won't be used again and so you won't see that it is not slow any more. You could periodically check all the servers to see if their performance has changed.