Re: Misaligned BufferDescriptors causing major performance problems on AMD

On Wed, Feb 5, 2014 at 11:37 AM, Andres Freund <andres(at)2ndquadrant(dot)com> wrote:
> On 2014-02-05 11:23:29 -0500, Tom Lane wrote:
>> Andres Freund <andres(at)2ndquadrant(dot)com> writes:
>> > And I think somebody already thought about it (c.f. ALIGNOF_BUFFER), it
>> > just wasn't updated in the last 10 years.
>>
>> No, ALIGNOF_BUFFER is there because we read something that said that I/O
>> transfers between userspace and kernel disk cache would be faster with
>> aligned buffers. There's been no particular thought given to alignment
>> of other data structures, AFAIR.
>
> But it's not aligned anymore on at last half a decade's hardware, and
> it's what we already align *all* bigger ShmemAlloc() values with. And
> BufferDescriptors surely counts as larger in its entirety.
>
>> It may well be that your proposal is spot on. But I'd like to see some
>> data-structure-by-data-structure measurements, rather than assuming that
>> alignment must be a good thing.
>
> I am fine with just aligning BufferDescriptors properly. That has
> clearly shown massive improvements.

I thought your previous idea of increasing BUFFERALIGN to 64 bytes had
a lot to recommend it. But that doesn't mean it doesn't need testing.

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