openmeetings-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "seba.wagner@gmail.com" <seba.wag...@gmail.com>
Subject Re: [DISCUSS] Manual vs Dynamic (Software Based) Load Balancing
Date Tue, 27 Nov 2012 11:47:01 GMT
Hallo Marcus,

please review this graphic:
https://cwiki.apache.org/confluence/display/OPENMEETINGS/Cluster+Master-Slave+overview

What we want to cluster is the RTMP traffic. That has nothing todo with
Tomcat 7, except that Red5 runs in the Tomcat 7 servlet container.

And the session that we are talking about are not Tomcat sessions, they are
RTMP-Connections. Each RTMP connection is a session. Those RTMP connections
are handled via Red5 and not via Tomcat. We have on top another session
object that has a 1:1 relation to a RTMP Connection.

Your solution might fit for the clustering of the Wicket part that is
indeed implemented using standard Tomcat technologies. However, we are
currently concentrating on RTMP traffic, cause in our current setup, this
is the bottleneck.

Sebastian


2012/11/27 Marcus Wellnitz <wellnitz@datenwerk-it.de>

> Hello,
>
> IMHO it would be fine to connect all cluster nodes via multicast (as
> tomcat does it for session replication mechanism). Multicast was designed
> for streaming data and in this case we HAVE streaming data.
>
> As I understand the OM architecture the first conference-user opens the
> room-session. At this OM-Instance the steam-management has to be done. All
> other client-streames can open an multicast-stream for their data and can
> connect to the mixed(master) mc-stream of the room.
>
> This setup is flexible and if you create a (tomcat) ajp adapter an apache
> reverse proxy can get the system status (load, #sessions, etc.) directly
> via the ajp-connector. With this information apache can do all
> loadbalancing thinks.
>
> This solution will ensure that the complexity of the LB system will stay
> moderate and if the implementition is smart enough it doesn't matter if a
> cluster node crashes. Apache will migrate the session to an other node with
> maybe one or two seconds of pause.
>
> Please look at the simpleTcpCluster at the tomcat doku:
> http://tomcat.apache.org/**tomcat-7.0-doc/config/cluster.**html<http://tomcat.apache.org/tomcat-7.0-doc/config/cluster.html>
> http://tomcat.apache.org/**tomcat-7.0-doc/config/ajp.html<http://tomcat.apache.org/tomcat-7.0-doc/config/ajp.html>
>
> Greetings from Germany
>
> Marcus
>
> Am 27.11.2012 04:37, schrieb Maxim Solodovnik:
>
>  I believe room capacity check should also be added.
>> I mean in our configuration we have 10 rooms and 2 servers
>> 9 rooms has capacity of 1-10
>> 1 room has capacity of 150
>>
>> in such situation it is better to have room with capacity of 150 on
>> separate server.
>>
>>
>> On Tue, Nov 27, 2012 at 10:30 AM, seba.wagner@gmail.com <
>> seba.wagner@gmail.com> wrote:
>>
>>  I have a request to discuss regarding clustering.
>>>
>>> Status: The syncing of the sessions in the cluster from the slaves to the
>>> master is basically ready. So the master of the cluster has a light
>>> version
>>> of the slave's session object, and it is possible for the master to find
>>> out the load across the cluster. The master can also kick out users from
>>> a
>>> conference room that is hosted on any slave.
>>>
>>> The question is know: How do we calculate which conference room is
>>> assigned
>>> to which server?
>>> Basically there is no need to have a configuration value in the
>>> organization or conference room, that assigns the room to any server in
>>> the
>>> cluster.
>>> You can simply do that dynamically: *As soon as the first user enters the
>>> conference room, the cluster checks which server has free capacities and
>>> starts the conference room on that slave.* And then anybody joining the
>>> conference will be redirected to the same server.
>>>
>>> I basically like this idea, as it makes it more dynamic and it is likely
>>> that you will use your given resources better with such a solution then
>>> by
>>> manually assigning slave/servers to conference rooms (often nobody uses
>>> those rooms and the slave/server will be just a zombie server that is
>>> blocked but has nothing todo).
>>>
>>> What do you think?
>>>
>>> Sebastian
>>> --
>>> Sebastian Wagner
>>> https://twitter.com/#!/dead_**lock <https://twitter.com/#!/dead_lock>
>>> http://www.webbase-design.de
>>> http://www.wagner-sebastian.**com <http://www.wagner-sebastian.com>
>>> seba.wagner@gmail.com
>>>
>>>
>>
>>
>
> --
> Marcus Wellnitz
> Datenwerk GmbH
> Vogelsbergstr. 34
> 60316 Frankfurt am Main
>
> Telefon +49 (0) 69/9434086-0
> Fax     +49 (0) 69/9434086-9
> E-Mail  hosting@datenwerk-it.de
> Web     www.datenwerk-it.de
>
> Sitz: Frankfurt am Main, Amtsgericht  Frankfurt am Main HRB 55221
> Geschäftsführer: Michael Beck
>
>


-- 
Sebastian Wagner
https://twitter.com/#!/dead_lock
http://www.webbase-design.de
http://www.wagner-sebastian.com
seba.wagner@gmail.com

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message