<!DOCTYPE html PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html>
<head>
<meta content="text/html;charset=UTF-8" http-equiv="Content-Type">
</head>
<body bgcolor="#ffffff" text="#000000">
Filip Rembiałkowski wrote:
<blockquote
cite="mid:92869e660808180728q34a80678v5931745868f820b4@mail.gmail.com"
type="cite">
<pre wrap="">2008/8/18 Julien Garet <a class="moz-txt-link-rfc2396E" href="mailto:Julien.Garet@inria.fr"><Julien.Garet@inria.fr></a>:
</pre>
<blockquote type="cite">
<pre wrap="">Hi,
I've just setup a testing cluster of postgresql with pglcuster and there
is something and do understand with the perfs. My setup is quite simple, I
have two identical boxes with database and replicator under centos 5, it
works well. I have done some testing with pgbench and a third identical box
running postgres 8.3 without postgres as reference.
The number of tps is about 5 times less in the cluster compared to the
reference box, using all kind of queries of pgbench (1 select, 1 insert et 3
update), which is quite normal, due to the overhead of the cluster in
writing data. But when I use pgbench, only using select queries, the nodes
of the cluster are still 5 times slower than the reference postgresql
server. And I don't understand that as pgcluster isn't supposed to replicate
data-read queries between nodes, the data-read performances should be the
same as the reference. Are there any clues to this mysterious thing ?
</pre>
</blockquote>
<pre wrap=""><!---->
Hello
How many client connections are you using?
</pre>
</blockquote>
I am using 16 clients<br>
<blockquote
cite="mid:92869e660808180728q34a80678v5931745868f820b4@mail.gmail.com"
type="cite">
<pre wrap="">
Are you connecting to pglb? If you connect to one of cluster dbs there
will be no load balancing.
</pre>
</blockquote>
I am not, I use a cisco load balancer (as I believe load balancing is a
network matter, not a software thing ;-)) I have run pgbench on all the
nodes + the load balanced IP.<br>
<blockquote
cite="mid:92869e660808180728q34a80678v5931745868f820b4@mail.gmail.com"
type="cite">
<pre wrap="">
Can you email pgbench commandline and results here?
</pre>
</blockquote>
see attached document. node1 and node2 are the nodes of the cluster,
slb is the load balanced address (which forward requests to the two
nodes with a least connection algorithm), ref the reference postgresql
server.<br>
<blockquote
cite="mid:92869e660808180728q34a80678v5931745868f820b4@mail.gmail.com"
type="cite">
<pre wrap="">
regards
</pre>
</blockquote>
thanks for your help<br>
<br>
<pre class="moz-signature" cols="72">--
Cordialement,
Julien Garet
Service des Moyens Informatiques
INRIA - Lille Nord Europe
Parc Scientifique de la Haute Borne
40, avenue Halley
59650 Villeneuve d'Ascq
Tél : 03.59.57.78.28
</pre>
</body>
</html>