Hi,
We want to limit the impact of (deep-)scrubs on PGs that contain a lot
of OMAP data on our cluster. The osd_scrub_sleep helps to limit impact
on cluster performance, but also dramatically slows down the scrubbing
process obviously. So I'm looking for knobs to turn to safely increase
scrubbing efficiency while limiting impact on client performance. There
seem to be a few candidates:
- name: osd_scrub_max_preemptions
type: uint
level: advanced
desc: Set the maximum number of times we will preempt a deep scrub
due to a client
operation before blocking client IO to complete the scrub
default: 5
min: 0
max: 30
I'm tempted to bump this to 30. This might slow down scrubs, but should
allow them progress sooner or later.
- name: osd_deep_scrub_keys
type: int
level: advanced
desc: Number of keys to read from an object at a time during deep scrub
default: 1024
with_legacy: true
Would increasing this value an order of magnitude (or more) help when
reading a lot of OMAP data? I.e. is frequently requesting a batch of
keys an expensive operation? Or would this be counter productive
(spending more time checking the data, hence locking the operations for
a longer period at a time)?
Are there other tunables that might reduce scrub impact for these kind
of PGs (and that do not negatively impact PGs that do not contain OMAP
data)?
Thanks,
Gr. Stefan
_______________________________________________
Dev mailing list -- dev@xxxxxxx
To unsubscribe send an email to dev-leave@xxxxxxx
