[LU-1157] improve flock deadlock detection: hash of waiting flocks instead of list

• Status: Resolved
• Project: Lustre
• Fix Version/s: Lustre 2.3.0
• Type: Improvement
• Priority: Minor
• Reporter: Vitaly Fertman
• Assignee: WC Triage
• Resolution: Fixed

Comments

[Andreas Dilger]

Vitaly, please also provide some useful information about this change.

[Vitaly Fertman]

http://review.whamcloud.com/2240

[Vitaly Fertman]

the test:
3000 threads takes by 1 non-conflicnting locks, after that these 3000 threads takes the lock its right neighbour thread has taken, so 3000-th lock gets deadlocked on the 1st one. to check it is dead locked, we need to search the list 3000 times on the last enqueue. this enqueue is done 2000 times. 2000 enqueues (only the deadlocked one is accounted) + 2000*3000 deadlock detections takes for following, in seconds:
===vanilla===
exports: 1 deadlocks: 2000 locks: 3000 time: 148
exports: 2 deadlocks: 2000 locks: 3000 time: 148
exports: 4 deadlocks: 2000 locks: 3000 time: 148
exports: 8 deadlocks: 2000 locks: 3000 time: 148
===per-export locking & hash===
exports: 1 deadlocks: 2000 locks: 3000 time: 19.5
exports: 2 deadlocks: 2000 locks: 3000 time: 12
exports: 4 deadlocks: 2000 locks: 3000 time: 11.5
exports: 8 deadlocks: 2000 locks: 3000 time: 10

[Vitaly Fertman]

this one is done on top of LU-1156

[Peter Jones]

Landed for 2.3

[Nathan Rutman]

Xyratex-bug-id: MRP-385