EDIT

I can't see that we have objects of size in CDM real data

I enabled G1 GC logging as described in https://plumbr.io/handbook/gc-tuning-in-practice/other-examples/humongous-allocations the to search for humongous object allocations on the test server. A couple of minutes after rebooting, the server still is starting instances, G1 Humongous Allocation are being reported:

 213.066: [G1Ergonomics (Concurrent Cycles) request concurrent cycle initiation, reason: requested by GC cause, GC cause: G1 Humongous Allocation]
213.067: [GC pause (G1 Humongous Allocation) (young) (initial-mark) 213.067: [G1Ergonomics (CSet Construction) start choosing CSet, _pending_cards: 31247, predicted base time: 83.62 ms, remaining time: 116.38 ms, target pause time: 200.00 ms]
 220.232: [G1Ergonomics (Concurrent Cycles) request concurrent cycle initiation, reason: requested by GC cause, GC cause: G1 Humongous Allocation]
 220.416: [G1Ergonomics (Concurrent Cycles) do not request concurrent cycle initiation, reason: concurrent cycle already in progress, GC cause: G1 Humongous Allocation]
 374.271: [G1Ergonomics (Concurrent Cycles) request concurrent cycle initiation, reason: requested by GC cause, GC cause: G1 Humongous Allocation]
 374.324: [G1Ergonomics (Concurrent Cycles) do not request concurrent cycle initiation, reason: concurrent cycle already in progress, GC cause: G1 Humongous Allocation]
 385.183: [G1Ergonomics (Concurrent Cycles) request concurrent cycle initiation, reason: requested by GC cause, GC cause: G1 Humongous Allocation]
 385.183: [G1Ergonomics (Concurrent Cycles) request concurrent cycle initiation, reason: requested by GC cause, GC cause: G1 Humongous Allocation]
385.183: [GC pause (G1 Humongous Allocation) (young) (initial-mark) 385.183: [G1Ergonomics (CSet Construction) start choosing CSet, _pending_cards: 40507, predicted base time: 76.38 ms, remaining time: 123.62 ms, target pause time: 200.00 ms]
 392.099: [G1Ergonomics (Concurrent Cycles) request concurrent cycle initiation, reason: requested by GC cause, GC cause: G1 Humongous Allocation]
 392.219: [G1Ergonomics (Concurrent Cycles) do not request concurrent cycle initiation, reason: concurrent cycle already in progress, GC cause: G1 Humongous Allocation]

so there is strong evidence that we in deed have big sized objects!

Andreas Kohlbecker wrote:

this finding can be especially relevant for imports and exports, therefore we should examine I/O functionalities which have been reported to cause problems in the near past. Isn't it that Walter had Problems a couple of weeks ago?

Do you remember anything like this Andreas & Katja?

Yes the cdmlight export for larger subtrees or a whole classification can cause memory problems.