Re: mysterious dbench results

[Marcelo Tosatti <marcelo@xxxxxxxxxxxxxxxx>]

On Fri, 23 Feb 2001, Steve Lord wrote:

<snip>

> Other changes coming down the pipeline - probably next week will help too,
> we have better dbench and bonnie numbers on some internal code right now.

It looks we're allocating the page to hold the page pointers for the
cluster unconditionally in __pagebuf_write_full_page (my last message
talks a bit more about that kind of stuff). This page, as far as I can
see, is not used at all without kiobuf io. (Thomas is not using kiobuf IO
IIRC).