Re: Locally Managed Tablespaces ... again!!!

"Howard J. Rogers" <howardjr2000_at_yahoo.com.au> wrote:
> >
> > I think DBWR contention does inherently arise between a (OLTP) table
> > and it's index if they are on the same spindle.
> >
>
> Yup, if all else fails, ignore the evidence,

You have provided no evidence, so it can't be ignored.

> and just re-assert that
> which is demonstrably untrue.

Demonstrate away.

> > The RDBMS "knows" about them when it automatically changes one
> > to reflects changes in the other. DBWR doesn't know this, but it
> > has to deal with the aftermath.
> >
>
> You stated that DBWR would be more likely to cause contention doing
> writes for indexes and tables than for any other pair of segments. Since
> it doesn't know what it's writing, it can't go out of its way to create
> that contention specially for that pair.

Just because you don't go out of way to cause something doesn't mean that something can't happen. Quite the opposite, somethings naturally happen unless you go out of your way to avoid them.

If the DBWR is faced with writing many blocks to the same spindle, contention will arise naturally.

> So you'll have to invent some
> other mechanism to explain why this pairing is particularly prone to
> write contention. Merely asserting that the index blocks and table blocks
> will be flushed in the same DBWR cycle doesn't do it, I'm afraid.

It seems pretty prima-facie to me. If you know of something that operates here to prevent that contention, I'm all ears. But I'm no more inclined to accept your dogma than I am the old dogma.

>
> > >
> > > And as re-written, that statement is perfectly true. But that's an
> > > argument for lots of spindles, not specifically to house indexes in a
> > > separate tablespace for performance reasons.
> >
> > Right, I wouldn't argue to separate tablespace from indices for
> performance
> > reasons.
>
> But if you're going to assert that tables and indexes are especially
> prone to write contention, that's exactly what you should do.

Well, that's what I should to if you aren't already using RAID or some other striping, and DBWR is the major and unacceptable bottleneck in your system, and it's due to spindle contention. Yes, under those conditions, that is what I should do with my current understanding of how Oracle works.

> You can't
> have it both ways.

Sure I can. Exploring hypothetical cases is one of the best way to learn things, but I certainly don't confuse them with reality.

Xho

-- 
-------------------- http://NewsReader.Com/ --------------------
Usenet Newsgroup Service              New Rate! $9.95/Month 50GB