Re: How to avoid transaction ID wrap

On Wed, Jun 07, 2006 at 01:48:50PM +0200, Zdenek Kotala wrote:
> Koichi Suzuki wrote:
> >I've once proposed a patch for 64bit transaction ID, but this causes
> >some overhead to each tuple (XMIN and XMAX). Pgbench with 64bit
> >transaction ID has to pay about a couple of percent of performance.
> >If 64bit transaction ID is a reasonable fix, I've already posted this
> >patch. Anyone can apply this to later versions.
> Did you check performance on 32-bit or 64-bit systems and 64-bit binary
> version of PGSQL? I think that today is not problem to have 64-bit
> architecture and 64-bit ID should increase scalability of Postgres.

I doubt performance is the issue directly. Increasing the size of the
transaction counter would increase the size of narrow tables by maybe
30%. That's 30% more disk space and 30% more memory usage in some
places. Maybe at some point it'll be worth it, but right now I don't
think those commodoties are cheap enough to use like this for fairly
marginal benefits.

Beside, memory bandwidth hasn't grown anywhere enar as fast as memory
space, so it's always a good idea to use as little memory as possible.

Have a nice day,
--
Martijn van Oosterhout <kleptog(at)svana(dot)org> http://svana.org/kleptog/
> From each according to his ability. To each according to his ability to litigate.