When deleting entries from the distributed cache we do so in a background thread. Once the size limit of the distributed cache is reached all unused entries are deleted.
MAPREDUCE-2494 changes this so that entries are deleted in LRU order until the usage falls below a given threshold. In either of these cases we are periodically flooding a disk with delete requests which can slow down all IO operations to a drive. It would be better to be able to throttle this deletion so that it is spread out over a longer period of time. This jira is to add in this throttling.
On investigating it seems much simpler to backport MPAREDUCE-2494 to 20S before implementing this change rather then try to implement it without LRU deletion, because LRU goes a long way towards reducing the load on the disk anyways.