There was no "fix" here, but I'm filing this analysis here as an (immediately closed) JIRA issue for future ref:
 If we very carefully comb over  E// attached initial/akka.conf, we can see that it is missing this line which upstream has:
 
snapshot-store.local.class = "org.opendaylight.controller.cluster.persistence.LocalSnapshotStore"

Thereforefore they were still using the default akka.persistence.snapshot.local.LocalSnapshotStore (seen in the stack trace yesterdday, below) instead of ODL's custom, and optimzed, org.opendaylight.controller.cluster.persistence.LocalSnapshotStore! Either you "just forgot" that, or, more likely, that was made after Boron... see here for the full history of this story in CONTROLLER-1576 and:

* https://github.com/opendaylight/controller/blob/master/opendaylight/md-sal/sal-clustering-commons/src/main/java/org/opendaylight/controller/cluster/persistence/LocalSnapshotStore.java
* https://github.com/opendaylight/controller/commits/master/opendaylight/md-sal/sal-clustering-commons/src/main/java/org/opendaylight/controller/cluster/persistence/LocalSnapshotStore.java
* https://git.opendaylight.org/gerrit/#/q/topic:bug/7521

Before this fix with a custom LocalSnapshotStore, the default one from Akka would create (x3 ?) a HUGE byte[] instead of "streaming directly to the file like the new custom one, so the JVM was not able to park it mid-stream, as Akka's version may hold up the safepoint while serializing to a byte[]" (summarized from private email exchange with Tom).
 
They als ran ODL with a huge 24 GB heap, which does not help for GC, saying "[Dhiman] The reason for huge heap size is i) to tackle with huge memory allocation during PM counter’s serialization (jolokia)", and that problem was solved in Jolokia 2.5 years ago, see https://github.com/rhuss/jolokia/issues/391; "and ii) Humongous allocation for huge byte array creations before writing the snapshots" which was solved by CONTROLLER-1576 (above) 1.5 years ago, according to the JavaDoc on org.opendaylight.controller.cluster.persistence.LocalSnapshotStore.

PS: Earlier private email about this problem suspected mmp memory mapping files related issues from LevelDB (native or pure Java impl) blocking the world before GC, but that was a wrong initial conclusion; the snapshot (contrary to the journal) actually has absolutely nothing to do with LevelDB.