We represents a prototyping of bandwidth control strategy of recovery activities at per-osd granularity by utilizing dmclock, an algorithm that implements distributed quality of service. The benefit is that we're now be able to limit the bandwidth of recovery activities in any form, thereby greatly reducing their impact on our existing client-side services.
I am an active and experienced Ceph developer focusing on RADOS. I've talked about the "Per OSD Recovery Bandwidth Control Based on dmClock" strategy with Yan on Cephalocon Barcelona 2019.