In the default configuration, when a disk becomes saturated, its disk queue will gradually fill up. Then, the filestore queue will start to fill up. Up until this point, I/O would have been accepted as fast as the journal could accept it. As soon as the filestore queue fills up and/or the WBThrottle kicks in, I/O will suddenly be stopped until the queues fall back below the thresholds. This behavior will lead to large spikes and, most likely, periods of low performance, where other client requests will experience high latency.

In order to reduce the spikiness of filestore when the disks become saturated, there are some additonal configuration options that can be set to gradually throttle back operations as the filestore queue fills up, instead of bouncing around the hard limit.