Bruce Momjian <maillist@candle.pha.pa.us> writes:
>> Tom Lane wrote:
>>>> I've been doing some more backend profiling, and observe that in a large
>>>> SELECT from a table with lots of columns, nocachegetattr (the guts of
>>>> heap_getattr) is at the top of the list, accounting for about 15% of
>>>> runtime.
>>>>
>>>> The percentage would be lower in a table with fewer columns or no null
>>>> columns, but it still seems worth working on. (Besides, this case right
>>>> here is a real-world case for me.)
> nocachegetattr() computes all offsets, even offsets after the column you
> are requesting, to prevent future calls. You must have nulls or
> varlena's that is causing nocachegetattr to be called so many times.
> Is this true?
Right, this table has 38 columns, many of which can be NULL and several
of which are variable-size. So it's probably the worst-case scenario as
far as the cost of nocachegetattr is concerned. It looked to me like
the pre-computation aspect of nocachegetattr only works for tables where
all the tuples have the same physical layout, ie, no varlenas or nulls;
is that right?
> heap_getattr() certainly is called many times, and needs any
> optimization we can give it. I have done as much as I could. Perhaps
> there are more opportunities I missed.
I thought I had spotted a couple of possibilities for small improvements
of the code inside nocachegetattr, but it was awfully late by then so
I didn't try changing anything. I'll take another look.
regards, tom lane