Heikki Linnakangas <heikki.linnakangas@enterprisedb.com> writes:
> ... but we still haven't actually
> established that the WAL-logging is causing the performance degradation
> Zdenek observed.
Yeah, that's a good point. I did some simple performance testing on
bulk inserts and updates, and found that while they indeed tended to be
WALInsertLock heavy, the FSM traffic seemed to be only a small part of
it. Here are some xlog record type counts from a bulk update test:
686555 XLogInsert: rm 10 info 20 HEAP_UPDATE 89117 XLogInsert: rm 10 info 29 HEAP_UPDATE + bkp blk + removable
24526XLogInsert: rm 10 info 25 HEAP_UPDATE + bkp blk + removable 3199 XLogInsert: rm 10 info 2d HEAP_UPDATE + 2
bkpblks + removable 27676 XLogInsert: rm 7 info 00 FSM_SET_AVAIL 35 XLogInsert: rm 7 info 09 SET_AVAIL + bkp
blk+ removable
So either by record count or by volume, the FSM traffic doesn't seem to
be much. I wonder whether Zdenek knows what the xlog traffic is like
for his test in an unpatched database ...
regards, tom lane