Thread: Question regarding clock-sweep

Question regarding clock-sweep

From
Josh Berkus
Date:
Tom,

Now that I'm beginning serious performance testing of clock-sweep, I was going 
back through the lock discussion and am not sure what the patch that actually 
went in 3 weeks ago consisted of.   Is it clock-sweep with a used/unused bit 
or a counter?   How is it handling seq scans?

Oh, and incidentally, can I use the same database files for 8.0.2 and 8.1cvs 
3/10/05?

-- 
Josh Berkus
Aglio Database Solutions
San Francisco


Re: Question regarding clock-sweep

From
Neil Conway
Date:
Josh Berkus wrote:
> Oh, and incidentally, can I use the same database files for 8.0.2 and 8.1cvs 
> 3/10/05?

No.

-Neil



Re: Question regarding clock-sweep

From
Tom Lane
Date:
Josh Berkus <josh@agliodbs.com> writes:
> Now that I'm beginning serious performance testing of clock-sweep, I
> was going back through the lock discussion and am not sure what the
> patch that actually went in 3 weeks ago consisted of.  Is it
> clock-sweep with a used/unused bit or a counter?  How is it handling
> seq scans?

It's clock-sweep with a counter.  The counter increments on reference,
up to a small maximum value (BM_MAX_USAGE_COUNT in buf_internals.h),
and decrements when the clock hand passes over the buffer.  I'd be
interested to see trials with different values of BM_MAX_USAGE_COUNT
... I made it 5 to start with but that was a WAG.

There's not any special smarts for seqscans, but the counter should
handle that.

> Oh, and incidentally, can I use the same database files for 8.0.2 and 8.1cvs 
> 3/10/05?

Sorry, we forced initdb already several times...
        regards, tom lane


Re: Question regarding clock-sweep

From
Josh Berkus
Date:
Tom, Neil,

> > Oh, and incidentally, can I use the same database files for 8.0.2 and
> > 8.1cvs 3/10/05?
>
> Sorry, we forced initdb already several times...

D'oh.  Better see if I can get a second machine, then, since build of a 100G 
DBT3 database takes 20 hours.

-- 
Josh Berkus
Aglio Database Solutions
San Francisco


Re: Question regarding clock-sweep

From
Josh Berkus
Date:
Tom,

> There's not any special smarts for seqscans, but the counter should
> handle that.

So all pages start out with the same counter, except VACUUM tuples?

BTW, I found this paper on RDBMS memory management, which probably could have 
saved us some weeks of discussion:
http://www.cs.usask.ca/~wew036/research/comprehensive.pdf

-- 
Josh Berkus
Aglio Database Solutions
San Francisco