Re: [HACKERS] maintenance_work_mem and CREATE INDEX time

On Wed, Jul 24, 2013 at 3:20 AM, Jeff Janes <jeff.janes@gmail.com> wrote:
> On Mon, Jul 22, 2013 at 9:11 PM, Amit Langote <amitlangote09@gmail.com> wrote:
>> Hello,
>>
>> While understanding the effect of maintenance_work_mem on time taken
>> by CREATE INDEX, I observed that for the values of
>> maintenance_work_mem less than the value for which an internal sort is
>> performed, the time taken by CREATE INDEX increases as
>> maintenance_work_increases.
>>
>> My guess is that for all those values an external sort is chosen at
>> some point and larger the value of maintenance_work_mem, later the
>> switch to external sort would be made causing CREATE INDEX to take
>> longer. That is a smaller value of maintenance_work_mem would be
>> preferred for when external sort is performed anyway. Does that make
>> sense?
>
> The heap structure used in external sorts is cache-unfriendly.  The
> bigger the heap used, the more this unfriendliness becomes apparent.
> And the bigger maintenance_work_mem, the bigger the heap used.
>
> The bigger heap also means you have fewer "runs" to merge in the
> external sort.  However, as long as the number of runs still fits in
> the same number of merge passes, this is generally not a meaningful
> difference.

Does fewer runs mean more time (in whichever phase of external sort)?


--
Amit Langote