On Sun, Nov 28, 2010 at 7:15 PM, Tom Lane <tgl@sss.pgh.pa.us> wrote:
> Robert Haas <robertmhaas@gmail.com> writes:
>> One possible way to get a real speedup here would be to look for ways
>> to trim the number of catcaches.
>
> BTW, it's not going to help to remove catcaches that have a small
> initial size, as the pg_am cache certainly does. If the bucket zeroing
> cost is really something to minimize, it's only the caches with the
> largest nbuckets counts that are worth considering --- and we certainly
> can't remove those without penalty.
Yeah, very true. What's a bit frustrating about the whole thing is
that we spend a lot of time pulling data into the caches that's
basically static and never likely to change anywhere, ever. I bet the
number of people for whom <(int4, int4) has any non-standard
properties is somewhere between slim and none; and it might well be
the case that formrdesc() is faster than reading the relcache init
file, if we didn't need to worry about deviation from canonical. This
is even more frustrating in the hypothetical situation where a backend
can switch databases, because we have to blow away all of these cache
entries that are 99.9% likely to be basically identical in the old and
new databases.
The relation descriptors for pg_class and pg_attribute are examples of
things it would be nice to hardwire and never need to update. We are
really pretty much screwed if there is any meaningful deviation from
what is expected, but relpages, reltuples, and relfrozenxid - and
maybe relacl or reloptions - can legitimately vary between databases.
Maybe we could speed things up a bit if we got rid of the pg_attribute
entries for the system attributes (except OID).
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company