The first thing that we can do is to split out
the c2s from the server and make it a seperate
process/component. Once that is done, we can
replicate the c2s processes over multiple
machines. Each one would connect back to the
main server running the JSM and everything
should work just fine.
Round Robin will not provide true load
balancing since there is no mechanism in
place to check how many connections a
server has when it forwards a new connection
to it the next time it comes up in the
Round Robin.
This is doable right now. The 1.5 version of
jabberd has sought to break all of the pieces
out into seperate processes. This was not
done for farming specifically, but we will not
complain since the jadc2s component can handle
upwards of 10k users. (The 1.4.x series c2s
could only handle ~1024).
Currently there is a forked version of jadc2s
that works with the 1.4.2 server. It is located
in the jabberd14 CVS repository on JabberStudio.
The following example is running with two jadc2s
boxes, and one central jabberd box. To set this
up do the following:
Get the
source
for xdb_sql from
IDEALX
, build it, and setup the jabber.xml to use it.
This is a very important step. xdb_file,
the default xdb that comes with jabberd,
is limited to open file descriptors too.
You can play the same shell games that we
are going to play with jadc2s later, but
if you want a server that can handle
millions or users, then you need something
other than xdb_file. Enter xdb_sql, which
only uses one file descriptor to connect
to the mysql server.
For more information on how to configure
xdb_sql, please see the README that is
distributed in the release.
Build jadc2s and distribute the binaries
to the boxes where they will run.
Setup the main jabberd to accept the jadc2s
component connections. In the jabber.xml
config file, add the following XML:
Now you can run the main jabberd and get it
listening for the jadc2s to connect to
it.
Configure the jadc2s.xml on each box to
connect to the SM, where to listen, etc...
<!-- session manager configuration -->
<sm>
<!-- host and port of the SM -->
<host>localhost</host>
<port>5111</port>
<!-- shared secret, for authenticating us -->
<secret>secret</secret>
<!-- our ID, for authenticating us to the sm -->
<id>jadc2s</id>
<!-- how many times to try to connect to the sm (default: 5) -->
<retries>5</retries>
</sm>
<!-- local network settings -->
<local>
<!-- who we identify ourselves as. This should correspond to the -->
<!-- ID (host) that the session manager thinks it is. -->
<id>localhost</id>
<!-- IP address to bind to (default: 0.0.0.0) -->
<!-- <ip>0.0.0.0</ip> -->
<!-- port to bind to (default: 5222) -->
<port>5222</port>
<!-- SSL configuration -->
<!-- Specify the port to listen on, and the key file to use for -->
<!-- the certificate. -->
<!-- <port/> (default: 5223) -->
<!-- <pemfile/> (default: ./server.pem) -->
<!--
<ssl>
<port>5223</port>
<pemfile>./server.pem</pemfile>
</ssl>
-->
</local>
For more information on how to configure
jadc2s, please see the README in the jadc2s
source directory.
Open a shell where you can change file
system parameters (root usually) and execute
the following command:
bash$echo "24000" > /proc/sys/fs/file-max
This bumps the upper bound on the
number of allowed file descriptors
that can be open at one time.
Set the limit for the shell you are in
to use more than the default 1024 file
descriptors.
bash$ulimit -n 11000
Tell jadc2s how many file descriptors it
is allowed to use:
<!-- maximum number of file descriptors. Should be a prime -->
<!-- number. At least four will be used by jadc2s itself, -->
<!-- usually around six or seven (default: 1023) -->
<!-- For a list of prime numbers: -->
<!-- http://www.utm.edu/research/primes/lists/small/10000.txt -->
<max_fds>10007</max_fds>
All that's left is to run the server, and the
jadc2s processes. Everything should work fine.
If it doesn't, then PLEASE tell me at
<reatmon@jabber.org> so that I can
fix this document.
