Re: Tuplesort merge pre-reading

On Thu, Sep 22, 2016 at 4:17 AM, Heikki Linnakangas <hlinnaka(at)iki(dot)fi> wrote:
> On 09/22/2016 03:40 AM, Claudio Freire wrote:
>>
>> On Tue, Sep 20, 2016 at 3:34 PM, Claudio Freire <klaussfreire(at)gmail(dot)com>
>> wrote:
>>>
>>> The results seem all over the map. Some regressions seem significant
>>> (both in the amount of performance lost and their significance, since
>>> all 4 runs show a similar regression). The worst being "CREATE INDEX
>>> ix_lotsofitext_zz2ijw ON lotsofitext (z, z2, i, j, w);" with 4GB
>>> work_mem, which should be an in-memory sort, which makes it odd.
>>>
>>> I will re-run it overnight just in case to confirm the outcome.
>>
>>
>> A new run for "patched" gives better results, it seems it was some
>> kind of glitch in the run (maybe some cron decided to do something
>> while running those queries).
>>
>> Attached
>>
>> In essence, it doesn't look like it's harmfully affecting CPU
>> efficiency. Results seem neutral on the CPU front.
>
>
> Looking at the spreadsheet, there is a 40% slowdown in the "slow" "CREATE
> INDEX ix_lotsofitext_zz2ijw ON lotsofitext (z, z2, i, j, w);" test with 4GB
> of work_mem. I can't reproduce that on my laptop, though. Got any clue
> what's going on there?

It's not present in other runs, so I think it's a fluke the
spreadsheet isn't filtering out. Especially considering that one
should be a fully in-memory fast sort and thus unaffected by the
current patch (z and z2 being integers, IIRC, most comparisons should
be about comparing the first columns and thus rarely involve the big
strings).

I'll try to confirm that's the case though.