Re: Bug in new buffering GiST build code

On 05.06.2012 09:45, Alexander Korotkov wrote:
> On Tue, Jun 5, 2012 at 2:00 AM, Alexander Korotkov<aekorotkov(at)gmail(dot)com>wrote:
>
>> On Mon, May 28, 2012 at 1:46 AM, Alexander Korotkov<aekorotkov(at)gmail(dot)com>wrote:
>>
>>> On Sat, May 26, 2012 at 12:33 AM, Heikki Linnakangas<
>>> heikki(dot)linnakangas(at)enterprisedb(dot)com> wrote:
>>>
>>>> Alexander, do you still have the test environments and data lying around
>>>> that you used for GiST buffering testing last summer? Could you rerun some
>>>> of those tests with this patch?
>>>
>>> I think I can restore test environment and data. Will rerun tests soon.
>>
>> I rerun some of tests. There are index build times in seconds for old way
>> of parent refind and new way of it.
>>
>> old new
>> usnoa2 2385 2452
>> usnoa2_shuffled 8131 8055
>> uniform 8327 8359
>>
>> I thinks difference can be described by round error.
>
> Oh, I mean not "round" error, but "random". I.e. not exactly same state of
> shared buffers at index build start and so on.

Thanks. I was expecting a small performance gain from the new approach,
but I guess not. Oh well, that would've just been a bonus, the important
thing is that it now works.

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com