In CRM Fuse, string deduplication table fixup takes a significant amount of time (~25-30% of average gc pause time).
One interesting thing noticed is that only around half of the threads actually do work. This may be due to startup/synchronization taking so long for many threads.
This is also somewhat similar to the problem described in JDK-8040120.
One option to fix this is to merge string deduplication work with other parallel tasks in the "Other" group if possible.