On Tue, Jul 26, 2011 at 3:34 PM, Simon Riggs <simon@2ndquadrant.com> wrote:
> You might be right, but I think we have little knowledge of how some
> memory barrier code you haven't written yet effects performance on
> various architectures.
>
> A spinlock per backend would cache very nicely, now you mention it. So
> my money would be on the multiple copies.
Maybe so, but you can see from the numbers in my OP that the results
still leave something to be desired.
> It's not completely clear to me that updating N copies would be more
> expensive. Accessing N low contention copies rather than 1
> high-contention value might actually be a win.
Yeah, I haven't tested that approach.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company