Message-ID: <45462710.3000301@kaltenbrunner.cc>
Date: Mon, 30 Oct 2006 17:23:44 +0100
From: Stefan Kaltenbrunner <stefan@kaltenbrunner.cc>
User-Agent: Thunderbird 1.5.0.7 (X11/20060928)
MIME-Version: 1.0
To: Luke Lonergan <llonergan@greenplum.com>
CC: "Spiegelberg, Greg" <gspiegelberg@cranel.com>,
	Worky Workerson <worky.workerson@gmail.com>,
	Merlin Moncure <mmoncure@gmail.com>, pgsql-performance@postgresql.org
Subject: Re: Best COPY Performance
References: <C16B625D.5708%llonergan@greenplum.com>
In-Reply-To: <C16B625D.5708%llonergan@greenplum.com>
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit

Luke Lonergan wrote:
> Stefan,
> 
> On 10/30/06 8:57 AM, "Stefan Kaltenbrunner" <stefan@kaltenbrunner.cc> wrote:
> 
>>> We've found that there is an ultimate bottleneck at about 12-14MB/s despite
>>> having sequential write to disk speeds of 100s of MB/s.  I forget what the
>>> latest bottleneck was.
>> I have personally managed to load a bit less then 400k/s (5 int columns
>> no indexes) - on very fast disk hardware - at that point postgresql is
>> completely CPU bottlenecked (2,6Ghz Opteron).
> 
> 400,000 rows/s x 4 bytes/column x 5 columns/row = 8MB/s
> 
>> Using multiple processes to load the data will help to scale up to about
>>   900k/s (4 processes on 4 cores).

yes I did that about half a year ago as part of the CREATE INDEX on a 
1,8B row table thread on -hackers that resulted in some some the sorting 
improvements in 8.2.
I don't think there is much more possible in terms of import speed by 
using more cores (at least not when importing to the same table) - iirc 
I was at nearly 700k/s with two cores and 850k/s with 3 cores or such ...

> 
> 18MB/s?  Have you done this?  I've not seen this much of an improvement
> before by using multiple COPY processes to the same table.
> 
> Another question: how to measure MB/s - based on the input text file?  On
> the DBMS storage size?  We usually consider the input text file in the
> calculation of COPY rate.


yeah that is a good questions (and part of the reason why I cited the 
rows/sec number btw.)


Stefan