Feature #15835
closed
filestore: randomize split threshold
Added by Josh Durgin almost 8 years ago.
Updated about 6 years ago.
Category:
Performance/Resource Usage
Description
If the distribution of files is roughly even, many osds will reach the split threshold at the same time, causing them all to incur high latency as they all split directories at once.
A simple change that may mitigate this is to randomize the split threshold, similar to the randomized scrub threshold, so different osds split directories over a larger period of time.
- Source changed from other to Support
Hi! I am an undergrad student wishing to contribute to CEPH, and I would like to work on this issue. Please let me know.
Thanks,
Peng Chen
This one is more about performance testing, and at this point I think effort there is better spent on bluestore than filestore, and bluestore does not have internal splitting like this at all.
Perhaps you'd like to try http://tracker.ceph.com/issues/18629 ?
- Backport set to jewel, kraken
- Status changed from New to Fix Under Review
- Status changed from Fix Under Review to Resolved
- Backport deleted (
jewel, kraken)
Perf testing is not indicating much benefit, so I'd hold off on backporting this.
I spoke too soon, there is significantly improved latency and throughput in longer running tests on several osds.
- Category deleted (
OSD)
- Status changed from Resolved to Pending Backport
- Project changed from Ceph to RADOS
- Category set to Performance/Resource Usage
- Assignee set to Josh Durgin
- Status changed from Pending Backport to Resolved
Also available in: Atom
PDF