Don Baccus wrote:
>
> Zeugswetter Andreas SB SD wrote:
>
> > This is one of the main problems of the current optimizer which imho rather
> > aggressively chooses seq scans over index scans. During high load this does
> > not pay off.
>
> Bingo ... dragging huge tables through the buffer cache via a sequential
> scan guarantees that a) the next query sequentially scanning the same
> table will have to read every block again (if the table's longer than
> available PG and OS cache) b) on a high-concurrency system other queries
> end up doing extra I/O, too.
>
> Oracle partially mitigates the second effect by refusing to trash its
> entire buffer cache on any given sequential scan. Or so I've been told
> by people who know Oracle well. A repeat of the sequential scan will
> still have to reread the entire table but that's true anyway if the
> table's at least one block longer than available cache.
One radical way to get better-than-average cache behaviour in such
pathologigal casescases would be to discard a _random_ page instead of
LRU page (perhaps tuned to not not select from 1/N of pages on that are
MRU)
-------------
Hannu
