On Tue, Apr 5, 2016 at 5:45 PM, Andres Freund <andres@anarazel.de> wrote:
On 2016-04-05 17:36:49 +0300, Alexander Korotkov wrote: > Could the reason be that we're increasing concurrency for LWLock state > atomic variable by placing queue spinlock there?
Don't think so, it's the same cache-line either way.
Yes, it's very unlikely.
> But I wonder why this could happen during "pgbench -S", because it doesn't > seem to have high traffic of exclusive LWLocks.
Yea, that confuses me too. I suspect there's some mis-aligned datastructures somewhere. It's hard to investigate such things without access to hardware.
But it's quite easy to check if it is alignment issue. We can try your patch but without removing mutex from LWLock struct. If it's alignment issue, then TPS should become stable again.