Hmm, these figures seem perfectly reasonable. It seems I was also wrong about the db-cache heap region not being shared between graphs. So you have to gather more information about was is going wrong. Some ideas:

1. does restarting nginx make a difference? Communication between JanusGraph server and gremlin clients runs over websockets. Make sure there is no issue with communication channels kept occupied.
2. https://lists.lfaidata.foundation/g/janusgraph-users/topic/79935654#2886 This thread gives an nginx setup using a DNS virtual IP address. Possibly this makes a difference in websocket communication issues.
3. disable the db-cache and see if this make a difference (this can only make a difference if there are many clients with transaction caches)
4. check the logs of janusgraph server for other indications

Best wishes,      Marc