Re: bad estimation together with large work_mem generates terrible slow hash joins

On Wed, Sep 10, 2014 at 3:02 PM, Tomas Vondra <tv(at)fuzzy(dot)cz> wrote:
> On 10.9.2014 20:31, Robert Haas wrote:
>> On Wed, Sep 10, 2014 at 2:25 PM, Heikki Linnakangas
>> <hlinnakangas(at)vmware(dot)com> wrote:
>>> The dense-alloc-v5.patch looks good to me. I have committed that with minor
>>> cleanup (more comments below). I have not looked at the second patch.
>>
>> Gah. I was in the middle of doing this. Sigh.
>>
>>>> * the chunks size is 32kB (instead of 16kB), and we're using 1/4
>>>> threshold for 'oversized' items
>>>>
>>>> We need the threshold to be >=8kB, to trigger the special case
>>>> within AllocSet. The 1/4 rule is consistent with ALLOC_CHUNK_FRACTION.
>>>
>>> Should we care about the fact that if there are only a few tuples, we will
>>> nevertheless waste 32kB of memory for the chunk? I guess not, but I thought
>>> I'd mention it. The smallest allowed value for work_mem is 64kB.
>>
>> I think we should change the threshold here to 1/8th. The worst case
>> memory wastage as-is ~32k/5 > 6k.
>
> So you'd lower the threshold to 4kB? That may lower the wastage in the
> chunks, but palloc will actually allocate 8kB anyway, wasting up to
> additional 4kB. So I don't see how lowering the threshold to 1/8th
> improves the situation ...

Ah, OK. Well, never mind then. :-)

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company