Re: [SSI-devel] LVS and bind()

[Bruce W. <br...@ka...>]

> > Would it then be trivial to just say if ip is CVIP or wildcard, 
> > register with lvs?
> > I'm sure there are a few cases, like when the port is ephimeral.  
> > The normal listening
> > case is  bind(wildcard,specificport), which would then automatically 
> > be listening on the CVIP(s).  
> > 
> > I would think this also would mean that we wouldn't have to configure 
> > the lvs redirector for specific ports - it would just redirect ports 
> > that binds() had registered with it?
> 
> 
> I guess what you want is if IP is INADDR_ANY and port is specified
> register with CVIP. But in this case which CVIP ? . As of now in the
> kernel space there is no mapping which says this node is real server for
> a particular CVIP. But I think this is a nice feature. I will look into
> this. 

I think the registration with CVIP should happen at listen time and
not bind.  That way we can make sure it is a listen; bind is used for
both listens and connects.  

Which CVIP?  All of them.  
Is this node a "real server" for a particular CVIP?  I think all servers
in the cluster should be real servers for any port they listen to.
Another thing we don't have to configure.


> 
> 
> Allowing this feature for outgoing connection will break the load
> balancing functionality and will prevent the  daemons  like apache
> running on  multiple nodes. That's because to allow outgoing connection
> to register itself with LVS i will have to go for a Clusterwide port
> space that breaks load balancing. But we could make it a /proc/
> configurable item which indicate this configuration is not interested in
> load balancing. 

To avoid this, we should consider registration at listen time for the
incoming connections.  For outgoing connections that we want to 
associate with a CVIP, we will have to maintain the clusterwide ephimeral 
port space.  However, this will not break load balancing for 
incoming connections.

bruce
> 
> -aneesh 
> 
> 
> 
> 
> -------------------------------------------------------
> This SF.NET email is sponsored by:
> SourceForge Enterprise Edition + IBM + LinuxWorld = Something 2 See!
> http://www.vasoftware.com
> _______________________________________________
> ssic-linux-devel mailing list
> ssi...@li...
> https://lists.sourceforge.net/lists/listinfo/ssic-linux-devel