On Sat, Nov 17, 2012 at 05:07:18PM +0100, Andres Freund wrote:
> > > I agree that tripling FOR SHARE cost is risky. Where is the added cost
> > > concentrated? Perchance that multiple belies optimization opportunities.
> >
> > Good question, let me play a bit.
>
> Ok, I benchmarked around and from what I see there is no single easy
> target.
> The biggest culprits I could find are:
> 1. higher amount of XLogInsert calls per transaction (visible
> in pgbench -t instead of -T mode while watching the WAL volume)
> 2. Memory allocations in GetMultiXactIdMembers
> 3. Memory allocations in mXactCachePut
> a) cache entry itself
> b) the cache context
> 4. More lwlocks acquisitions
>
> We can possibly optimize a bit with 2) by using a static buffer for
> common member sizes, but thats not going to buy us too much...
In that case, +1 for your proposal to prop up FOR SHARE.