Feature #4930
open
OSD: replica to primary back pressure
Added by Samuel Just almost 11 years ago.
Updated about 5 years ago.
Description
Currently, replicas cannot throttle sub-ops. This means that a slower OSD can have OSDOps starved by sub-ops.
- Target version deleted (
v0.63)
This can probably be done in a few steps:
1 Add OSDService::estimate_delay() which returns an estimate of latency for a small write.
2 expose above via admin_socket and osd_stat for visibility
3 piggy-back the current delay on MOSDSubOp and MOSDPing as a way of allowing each OSD to build maps of peer delay estimates.
4 Rate limit sub-ops sent to any single at a rate which is a function of the current delay
- the goal is keep the peer saturated, but back off once latency starts to increase
1 & 2 are useful for visibility purposes independent of the overall task.
- Project changed from Ceph to RADOS
- Category deleted (
OSD)
- Component(RADOS) OSD added
Also available in: Atom
PDF