|
Hello,
On Fri, 16 Mar 2001, Steve wrote:
> At 01:12 16/03/2001 +0000, you wrote:
>
> > What do you want exactly? To make sure the indirect queries
> >always work (by providing high availability for the indirect
> >queries)? May be if you hit the VIP:177 with these indirect queries
> >you can then select the real X client. May be you can -broadcast
> >to all hosts in the LAN without using LVS? You can even use software
> >that can use a list with hosts to receive the broadcast message from
> >your X servers, in your case the hosts you try to use as XDMCP
> >real servers. I assume these broadcast XDMCP capabilities can help you
> >without using LVS?
> >
> > > Please note that I do have a box up and running that I can connect
> > > directly
> > > to.. It's just clustering it that I have the problem with :)
> >
> > I assume you have more hosts :) What kind of problems?
>
> The real-servers in the cluster are very powerful boxes (well compared with
> the lab machines anyway) as all three real-servers are dual Pentium III -
> 800Mhz with 512MB ram and Ultra-wide SCSI hardisks (nice toys!)
>
> In the computer lab we run 50% linux 50% windows2000, The aim of putting
> XDMCP onto the cluster would be to take advantage of the nice fast boxes to
> provide remote X sessions to the people using the windows machines. I would
> like it clustered as we already have the cluster, and it is powerful enough
> to deal with the extra load.
>
> Additionally we can then provide a shortcut on the desktop of the windows
> machines which pops up the XDMCP machine selection window (using exceed)
> and people can just click on the only machine there, and get a connection.
> (otherwise the poor undergraduates will all get confused, or all use the
> top box which will probably be some pathetic P2-233 and it will die
> horribly from the extra load)
>
> The real problem is more getting X sessions balanced across the servers
> (I'm not even sure if this is possible), as if there wasn't XDMCP we could
> always create a X session that only connected to the cluster.
May be you need a new chooser (probably guarded from LVS) that
will allow the client to select only one X Windows client host after
deciding which is the least loaded one. The way XDMCP is working allows
the first who replies for broadcast to serve the X session for the
X server. But I have never tried balancing XDMCP. May be there is
something already invented. If you want LVS to work like a chooser
may be someone must write such scheduler that will reply the indirect
queries and will allow the X server to select the least loaded X
client that is willing to manage the session. May be the health checking
software must perform the forward queries to the real servers to
determine whether they are willing to manage. Not sure, may be the
things can be more complex because it needs talking UDP but LVS does
not generate UDP packets/replies.
> This whole idea is to try and cater for the wider problem which is whatever
> the operating system balance in the lab, people always complain that there
> are either not enough linux or not enough windows machines :).
>
>
> > > Thanks,
> > >
> > >
> > > Steve.
>
> Steve.
>
> ----------------------------------------------------------------------------
> Before you criticize someone, walk a mile in his shoes. That way, if he gets
> angry, he'll be a mile away - and barefoot.
Regards
--
Julian Anastasov <ja@xxxxxx>
|
