On Sat, Nov 19, 2011 at 12:33 PM, Stephen Frost <sfrost@snowman.net> wrote:
> You've mentioned that before and, to be honest, I could have sworn that
> we're doing that already..
I tried to write a patch for that at one point, but it crashed and
burned over the exact same set of issues discussed upthread, which I
wasn't able to resolve satisfactorily. It's just really difficult to
change the API for something like memory allocation after the fact;
it's too hard to find the bits of code that do whatever naughty thing
you don't want them to.
One random idea... would there by any sense in having a palloc-like
function that is defined to allocate multiple objects at once? In
other words, if you need 4 list cells, then instead of asking palloc
for them one at a time, you make one function call and get four
pointers back at one go. I'm not sure whether that would help at all;
palloc might not be able to take advantage of the additional
information usefully.
To some extent I feel like this is all optimizing something that's
likely already so well-optimized that future gains, if any, are likely
to be small. I feel like the only way we're likely to get much of a
win here is if we can reduce the amount of memory that has to be
allocated in the first place (allocate fewer data structures, don't
copy them as much, etc.).
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company