From: Tom Lane <tgl@sss.pgh.pa.us>
To: Gurpreet Aulakh <gaulakh@ecmarket.com>
Cc: pgsql-performance@postgresql.org
Subject: Re: Query slower on 8.0.3 (Windows) vs 7.3 (cygwin)
Date: Fri, 23 Sep 2005 17:12:49 -0400
Message-ID: <6518.1127509969@sss.pgh.pa.us> (raw)
In-Reply-To: <IGEGINLEEDLANNHAAKCMIEDCCCAA.gaulakh@ecmarket.com>
References: <IGEGINLEEDLANNHAAKCMIEDCCCAA.gaulakh@ecmarket.com>
"Gurpreet Aulakh" <gaulakh@ecmarket.com> writes:
> After further investigation I have found that the reason why the query is
> slower on 8.0.3 is that the hash and hash joins are slower on the 8.0.3.
> So the question comes down to : Why are hash and hash joins slower?
I looked into this a bit and determined that the problem seems to have
been introduced here:
2002-12-30 10:21 tgl
* src/: backend/executor/nodeHash.c,
backend/executor/nodeHashjoin.c, backend/optimizer/path/costsize.c,
include/executor/nodeHash.h: Better solution to integer overflow
problem in hash batch-number computation: reduce the bucket number
mod nbatch. This changes the association between original bucket
numbers and batches, but that doesn't matter. Minor other cleanups
in hashjoin code to help centralize decisions.
(which means it's present in 7.4 as well as 8.0). The code now
groups tuples into hash batches according to
(hashvalue % totalbuckets) % nbatch
When a tuple that is not in the first batch is reloaded, it is placed
into a bucket according to
(hashvalue % nbuckets)
This means that if totalbuckets, nbatch, and nbuckets have a common
factor F, the buckets won't be evenly used; in fact, only one in every F
buckets will be used at all, the rest remaining empty. The ones that
are used accordingly will contain about F times more tuples than
intended. The slowdown comes from having to compare these extra tuples
against the outer-relation tuples.
7.3 uses a different algorithm for grouping tuples that avoids this
problem, but it has performance issues of its own (in particular, to
avoid integer overflow we have to limit the number of batches we can
have). So just reverting this patch doesn't seem very attractive.
The problem no longer exists in 8.1 because of rewrites undertaken for
another purpose, so I'm sort of tempted to do nothing. To fix this in
the back branches we'd have to develop new code that won't ever go into
CVS tip and thus will never get beta-tested. The risk of breaking
things seems higher than I'd like.
If we did want to fix it, my first idea is to increment nbatch looking
for a value that has no common factor with nbuckets.
regards, tom lane
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Reply to all the recipients using the --to and --cc options:
reply via email
To: pgsql-performance@postgresql.org
Cc: tgl@sss.pgh.pa.us, gaulakh@ecmarket.com
Subject: Re: Query slower on 8.0.3 (Windows) vs 7.3 (cygwin)
In-Reply-To: <6518.1127509969@sss.pgh.pa.us>
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
This inbox is served by DDX for PostgreSQL; see mirroring instructions
for how to clone and mirror all data and code used for this inbox