On Sat, Jul 11, 2015 at 2:28 AM, Jeff Davis <pgsql@j-davis.com> wrote:
> After talking with a few people at PGCon, small noisy differences in CPU
> timings can appear for almost any tweak to the code, and aren't
> necessarily cause for major concern.
I agree with that in general, but the concern is a lot bigger when the
function is something that is called everywhere and accounts for a
measurable percentage of our total CPU usage on almost any workload.
If memory allocation got slower because, say, you added some code to
regexp.c and it caused AllocSetAlloc to split a cache line where it
hadn't previously, I wouldn't be worried about that; the next patch,
like as not, will buy the cost back again. But here you really are
adding code to a hot path.
tuplesort.c does its own accounting, and TBH that seems like the right
thing to do here, too. The difficulty is, I think, that some
transition functions use an internal data type for the transition
state, which might not be a single palloc'd chunk. But since we can't
spill those aggregates to disk *anyway*, that doesn't really matter.
If the transition is a varlena or a fixed-length type, we can know how
much space it's consuming without hooking into the memory context
framework.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company
