To: Mark Kirkwood <markir@paradise.net.nz>
Cc: Luke Lonergan <llonergan@greenplum.com>,  stange@rentec.com,
	Greg Stark <gsstark@mit.edu>,  Dave Cramer <pg@fastcrypt.com>,
	Joshua Marsh <icub3d@gmail.com>,  pgsql-performance@postgresql.org
Subject: Re: Hardware/OS recommendations for large databases (
References: <BFAAB702.1460D%llonergan@greenplum.com>
	<438583C8.4050809@paradise.net.nz>
In-Reply-To: <438583C8.4050809@paradise.net.nz>
From: Greg Stark <gsstark@mit.edu>
Organization: The Emacs Conspiracy; member since 1992
Date: 24 Nov 2005 11:00:25 -0500
Message-ID: <87d5kqkpd2.fsf@stark.xeocode.com>
Lines: 24
User-Agent: Gnus/5.09 (Gnus v5.9.0) Emacs/21.4
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii

Mark Kirkwood <markir@paradise.net.nz> writes:

> Yeah - it's pretty clear that the count aggregate is fairly expensive wrt cpu -
> However, I am not sure if all agg nodes suffer this way (guess we could try a
> trivial aggregate that does nothing for all tuples bar the last and just
> reports the final value it sees).

As you mention count(*) and count(1) are the same thing.

Last I heard the reason count(*) was so expensive was because its state
variable was a bigint. That means it doesn't fit in a Datum and has to be
alloced and stored as a pointer. And because of the Aggregate API that means
it has to be allocated and freed for every tuple processed.

There was some talk of having a special case API for count(*) and maybe
sum(...) to avoid having to do this.

There was also some talk of making Datum 8 bytes wide on platforms where that
was natural (I guess AMD64, Sparc64, Alpha, Itanic).

Afaik none of these items have happened but I don't know for sure.

-- 
greg