I wrote:
> Gregory Stark <stark@enterprisedb.com> writes:
>> Is there any reason to experiment with this? I would have thought we would
>> divorce TOAST_MAX_CHUNK_SIZE from TOAST_THRESHOLD and hard code it as the same
>> expression that's there now. Ie, the largest size that can fit in a page.
> No, right now it's the largest size that you can fit 4 on a page. It's
> not obvious to me that 4 is optimal once it's divorced from TOAST_THRESHOLD.
> It seems possible that the correct number is 1, and even if it's useful
> to keep the tuples smaller than that, there's no reason to assume 4 is
> the best number per page.
I've just committed changes that make it trivial to experiment with the
number of toast tuples per page:
#define EXTERN_TUPLES_PER_PAGE 4 /* tweak only this */
/* Note: sizeof(PageHeaderData) includes the first ItemId on the page */
#define EXTERN_TUPLE_MAX_SIZE \ MAXALIGN_DOWN((BLCKSZ - \ MAXALIGN(sizeof(PageHeaderData) +
(EXTERN_TUPLES_PER_PAGE-1)* sizeof(ItemIdData))) \ / EXTERN_TUPLES_PER_PAGE)
#define TOAST_MAX_CHUNK_SIZE \ (EXTERN_TUPLE_MAX_SIZE - \
MAXALIGN(offsetof(HeapTupleHeaderData,t_bits)) - \ sizeof(Oid) - \
sizeof(int32)- \ VARHDRSZ)
Anyone who's got time to run performance experiments, have at it ...
regards, tom lane