I'm looking at a Java Flight Recording obtained from (internal) scale lab testing, , based on Oxygen SR3 code, and am surprised to find that the biggest "TLAB allocation", with a huge factor of several magnitudes (4.97 TiB vs. next issue at 86/57/16/15 ... GiB!), is this:

Iterator java.util.concurrent.ConcurrentHashMap$EntrySetView.iterator()	127113
void org.opendaylight.openflowplugin.applications.frm.nodeconfigurator.NodeConfiguratorImpl$JobQueueHandler.run()	127113
void java.lang.Thread.run()	127113

I'm not clear how object creation churn, and thus required GC, for a simple iterator() could be reduced, but it seems worth having a closer look at this code.