Thread: Re: pgbench questions

Re: pgbench questions

From
Tatsuo Ishii
Date:
[Cc:ed to hackers]

From: nconway@klamath.dyndns.org (Neil Conway)
Subject: pgbench questions
Date: Sat, 13 Jul 2002 00:57:37 -0400
Message-ID: <20020713045736.GA9258@klamath.dyndns.org>

> Hi,
> 
> I was looking at doing some performance profiling on PostgreSQL, and
> I had a few questions on pgbench.
> 
> (1) Is there a reason you chose to use the TPC-B benchmark rather
>     than TPC-C or TPC-H, for example? Do you think there might be any
>     merit in converting pgbench to use TPC-H or AS3AP?

Just easy to implement. Ideally pgbench should be able handle to
several kinds of benchmarks (I don't have time to do that sigh...)

BTW, TPC-H is very different from othe benchmarks. As far as I know,
it focuses on Data Ware House. So TPC-H cannot be a replacement for
TPC-B.

> (2) At least in the current CVS version, the code to do a 'CHECKPOINT'
>     after creating a table has been #ifdef'ed out. Why is that?

That is not after creation of a table, but while creating it, which is
not necessary any more since Tom has fix the growth of WAL logs.

> (3) Several people (Rod Taylor, Tom Lane, myself, perhaps others) have
>     noticed that the results obtained from pgbench can be somewhat
>     inconsistent (i.e. can vary between runs quite a bit).
> 
>     Have you found this to be the case in your own experience?
> 
>     Do you have any suggestions on how pgbench can be made more
>     consistent (either through special benchmarking procedures, or
>     through a change to pgbench)

I believe it's a common problem with benchmark programs. I think Tom
or Jan has posted a good summary to hackers list showing how to get
a consistent result with pgbench.
--
Tatsuo Ishii


Re: pgbench questions

From
"J. R. Nield"
Date:
On Sat, 2002-07-13 at 02:25, Tatsuo Ishii wrote: 
> > (2) At least in the current CVS version, the code to do a 'CHECKPOINT'
> >     after creating a table has been #ifdef'ed out. Why is that?
> 
> That is not after creation of a table, but while creating it, which is
> not necessary any more since Tom has fix the growth of WAL logs.
> 

Tatsou:

Could you or Tom give me some background on what this change was about?
Is this something recent, or would it have been in CVS about a month
ago?

I can't see any reason to force a checkpoint after CREATE TABLE, but it
would be interesting to know why it was done before.

;John Nield

-- 
J. R. Nield
jrnield@usol.com





Re: pgbench questions

From
Bruce Momjian
Date:
J. R. Nield wrote:
> On Sat, 2002-07-13 at 02:25, Tatsuo Ishii wrote: 
> > > (2) At least in the current CVS version, the code to do a 'CHECKPOINT'
> > >     after creating a table has been #ifdef'ed out. Why is that?
> > 
> > That is not after creation of a table, but while creating it, which is
> > not necessary any more since Tom has fix the growth of WAL logs.
> > 
> 
> Tatsou:
> 
> Could you or Tom give me some background on what this change was about?
> Is this something recent, or would it have been in CVS about a month
> ago?
> 
> I can't see any reason to force a checkpoint after CREATE TABLE, but it
> would be interesting to know why it was done before.

7.1.0 had a problem with WAL file growth.  This was fixed in 7.1.2 or
7.1.3 so checkpoint is not required anymore.

--  Bruce Momjian                        |  http://candle.pha.pa.us pgman@candle.pha.pa.us               |  (610)
853-3000+  If your life is a hard drive,     |  830 Blythe Avenue +  Christ can be your backup.        |  Drexel Hill,
Pennsylvania19026