Re: [WIP] Reduce alignment requirements on 64-bit systems.

Hello Zdenek,

On Thu, Oct 9, 2008 at 12:38 AM, Zdenek Kotala <Zdenek(dot)Kotala(at)sun(dot)com> wrote:
> Ryan Bradetich napsal(a):
>> On Wed, Oct 8, 2008 at 10:59 PM, Zdenek Kotala <Zdenek(dot)Kotala(at)sun(dot)com>
>> wrote:

>> I would like to see this patch (or some variant) go in if possible.
>> Since the inplace
>> upgrades a concern to you, is there anything I can do to help with the
>> inplace
>> upgrades to help offset the disruption this patch causes you?
>
> Yaah, wait until 8.5 :-). However, currently there is no clear consensus
> which upgrade method is best. I hope that It will clear after Prato
> developers meeting. Until this meeting I cannot say more.

Heh, understood. :)

I believe the proposed CRC patch also affects the heap page layout (adds
the pd_checksum field to the PageHeaderData). [1] Just pointing out another
patch that could affect you as well. My offer to help still stands.

> I overlooked 'd' test. Your idea seems to me reasonable. Maybe, you could
> test 'd' alignment only for NOT NULL values.

Funny you should mention this. I had just started looking into this
optimization
to see if I could convince myself it would be safe. My initial
conclusion indicates
it would be safe, but I have not tested nor verified that yet. Having
an independent
proposal for this boosts my confidence even more. Thanks!

> The problem there is add_item which it is used for indexes as well and they
> has IndexTupleHeader structure. I'm not convenience about idea has two
> different alignment for items on page.

Just to clarify, this patch only affects heap storage when (i.e. the
is_heap flag is
set). I have not had a chance to analyze or see if I can reduce
other storage types
yet.

> I guess another problem is with MAX_TUPLE_CHUNK_SIZE which uses MAXALIGN for
> computing. It seems to me that toast chunk could waste a space anyway.
>
> And of course you should bump page layout version.

Thanks. I will do.

> I also suggest create function/macro to compute hoff and replace code with
> this function/macro.

Great. That is some of the feedback I was looking for. I did not
implement it yet,
because I wanted to see if the basic implementation was feasible first.

Thanks again for your feedback!

- Ryan

[1] http://archives.postgresql.org/pgsql-hackers/2008-10/msg00070.php