Bug #459
closedbonnie++ is slow on clustered mds
100%
Description
We tracked it down to a problem with cap revocation while deleting inodes. The MDS is requesting that the kclient drop all caps, but the kclient isn't responding right away so there's a ~5-second wait on each file deletion.
This looks like it may only occur on inodes that have migrated auth from one MDS to another.
Need to re-run with kernel logging on to find out what's going on, we think we've tracked it down to being an issue there rather than on the MDS side. Or it may be on both sides, who knows.
Updated by Greg Farnum over 13 years ago
- Status changed from New to Resolved
- % Done changed from 0 to 100
Solved the most apparent issue, which is that if the kclient had already dropped caps for the MDS on an existing inode, it didn't rush that drop message back.
Then fixed a problem where the MDS is constantly bumping locks to sync, despite the fact that it looks like it shouldn't be. Because it was accidentally comparing the inode state instead of the lock state. Fixed the typo, and most deletes are taking more like .15 seconds instead of .7 seconds.
Pushed in commit:217fbdc9e422185bee09a1630e3343cf8baa20a3