Details
-
Improvement
-
Resolution: Unresolved
-
Critical
-
None
-
None
-
None
-
3
-
9223372036854775807
Description
While worked on unlink and zfs scalability problems
I found object free eats so much time of unlink, it caused an extra cpu load in the ARC cache flush thread.
16) <...>-2682981 | .... 7.184 us | mdd_trans_stop [mdd](); 16) <...>-2682981 | ....+ 62.367 us | } /* mdd_unlink [mdd] */ 16) <...>-2682981 | .... 0.030 us | mutex_unlock(); 16) <...>-2682981 | ....+ 43.762 us | lu_object_free.isra.36 [obdclass](); 16) <...>-2682981 | ....+ 46.206 us | } /* lu_object_put [obdclass] */ 16) <...>-2682981 | ....! 154.269 us | } /* mdt_reint_unlink [mdt] */
In short evaluation unlink perf increased by 10%-20% by moving object free into separate thread. ZFS have benefits from this change also.
"Alexey Lyashkov <alexey.lyashkov@hpe.com>" uploaded a new patch: https://review.whamcloud.com/c/fs/lustre-release/+/58910
Subject: LU-18938 obdclass: delay an object free
Project: fs/lustre-release
Branch: master
Current Patch Set: 1
Commit: d2ec441d7a71a816023dc5639d645da37df51176