Priority: Minor - P4
Affects Version/s: 3.0.1, 3.0.2, 3.0.3, 3.0.4, 3.0.5, 3.0.6
Fix Version/s: None
In some test case, Non periodic small stall is observed.
First time, I have thought wiredtiger checkpoint cause this small stall.
But, it's nothing to do with checkpoint (After some test and code eval, I found that wiredtiger table lock could cause this stall).
And this small stall makes a lot of slow query(over 150ms).
During this stall period, __wt_spin_lock(pthread_mutex_lock) took over 4~8 milliseconds.
__wt_schema_open_table :: elapsed 21 us
__wt_spin_lock :: elapsed:8394 us (pthread_mutex_lock :: session.open_cursor -> table creation(0))
__wt_schema_add_table :: elapsed 8436 us (need table lock)
__wt_schema_get_table :: elapsed 8437 us
__wt_curtable_open :: elapsed 8439 us
__wt_open_cursor :: elapsed 8456 us
My test case is really simple, Just make 1000 client thread and doing update(upsert:true) and total documents is about 400000.
And server have processed about 6~7000 update/sec. System load was under 1.0
12 core cpu, 32GB memory, 2.6.32-573.7.1.el6.centos.plus.x86_64
Looks like disk is nothing to do with this issue (I've tested NVME, SAS+RAID, SATA)
Current implementation, table lock is implemented with WT_SPINLOCK.
What do you think about changing it as WT_RWLOCK ?