Windows terminal server 2003 cluster

Question

We have a FOG with three Solaris 10 x86 servers, SRSS 4.2, SRWC 2.3 with the latest SRSS 4.2 patches. All Sunrays have the latest firmware.
Kiosk settings points to a windows 2003 terminal server cluster. The two nodes in the windows cluster have latest SRWC 2.0.
The cluster entitles, two terminal servers and a session manager server, we use roaming profiles.
When pointing kiosk settings to the cluster FQDN the sunray sessions are routed to just one server in the windows cluster. But if we RDP to Cluster's FQDN from a regular desktop the windows cluster load balances properly.
I do need help understanding how SRSS handles sessions? specifically if nodes are behind a cluster with session manager. Could it be that the session manager it's incompatible with SRSS in kiosk mode since it tries to route sessions at the mean time SRSS is doing the same task?
Thanks,
RE
Edited by: estrar on Dec 5, 2010 5:43 PM

Answer 1

Hi Estrar,
Short answer: 2003 Session Directory does not do load balancing.
Long answer (to the above and your other questions):
The Sun Ray Server doesn't do anything with regards to load balancing the windows sessions. To the Sun Ray Server, or more appropriately the windows connector, it's just making the connection you tell it to.
With regards to the windows client "working", am I correct in guessing that you ran subsequent instances of MS-RDC to "prove" that it worked?
Most likely what happened there is that the PC, thus MS-RDC, received different IP addresses back from DNS for the FQDN with each launch. This is basically round robin, not load balancing. But I'd have to understand how you've done the FQDN of the cluster in DNS.
The most common initial reaction is that the DNS server is setup incorrectly (I'll plead guilty that I've had the same reaction). It is true that some DNS servers can be configured will return the same IP on the first dns query. This is called a cyclic ordering and results in the IP address of the resolved name every client to be identical. For example, if a DNS server is set for cyclic ordering, every first request for the IP of server A for xyz.com will be the same on all clients. The second query from the same client should result in server B being resolved, if not, then you do have a DNS problem.. However, in the case of Sun Ray Server, the "second query from the same client" would be the second session coming up, since, as far as the DNS server is concerned, all the Sun Ray session are coming from the same client. Therefore, even with cyclic ordering, round robin should work even better in a Sun Ray environment then a "fat" client environment. You can check the documentation for specific DNS servers on how to set a random order for round robin (which should be the default), but chasing down cyclic ordering of RR entries as a cause to this "issue" is basically a red herring in a Sun Ray environment.
More than likely the same IP being returned everytime is a side effect of the name service caching daemon (nscd). The job of nscd is to speed up name resolution. However, ncsd will break any round robin scheme as it caches the first server returned by the first query from the caller (i.e. the host doing the name lookup, aka the Sun Ray server) and all applications on the caller will use that address for the lifetime of the cache. The default lifetime of the nscd cache is 3600 seconds (an hour). To disable the name service caching daemon, run the following command on every Sun Ray server in the host group (aka FOG): svcadm disable system/name-service-cache
An über smart colleague mentioned the possibility of another problem that could exist even if nscd is disabled and the hosts are on the same subnet as the caller . This has to do with default behavior of the Solaris resolver library putting servers that reside on the same subnet as the caller at the top of the sort order. You can read the notes in the man page for "gethostbyname" and also the man page for nss(4) for more information, but both reference how this behavior can break round robin. The fix is to edit /etc/default/nss and uncomment (or add) the statement SORT_ADDRS=FALSE
While those two things will allow a Solaris based Sun Ray server (or any Solaris client) to properly recognize round robin entries, it still holds true that 2003 Session Directory does not do load balancing. The only role of session directory is to ensure that you don't get routed to a different terminal server in case you have an existing session on another server. Round Robin + Session Directory may be all you need, but bear in mind that it is not "load balancing". Users could still connect directly to one of the IP addresses vs the round robin DNS name and that would defeat the round robin scheme in place.
Load balancing is a separate component in 2003, either using an external device like F5's BigIP or using Windows Network Load Balancing (NLB) service.
The type of load balancing solution that can be used depends on what is meant when you use the term "cluster". If the cluster is a real implementation of Windows 2003 Clustering (i.e.the type of cluster you'd use for SQL, Exchange, etc) then you can't use NLB and would have to use an external load balancer. Sun Ray Windows Connector supports either IP based or token based load balancers, though we don't certify any third party products in this regard.
"Cluster confusion" occurs because Session Directory itself is supported by Windows 2003 Clustering in order to provide a highly available directory service. Easy to see how this would be desirable with a large number of terminal servers. However, NLB does not work, nor is it supported on Windows 2003 Cluster.
To further the confusion, when NLB is enabled on the NIC that is handling RDP traffic, one of the configuration tabs is "Cluster Parameters". But it is important to realize that this has nothing to do with Windows 2003 Clustering and is in fact incompatible with it. NLB makes use of independent servers (i.e. not clustered), which the primary requirement of those servers is that they are on the same subnet.
Finally, it is easy to get confused between 2003 and 2008 session directory when it comes to load balancing, but they are very different. 2008 "session directory" does do load balancing. In fact, to underscore this distinction/feature, Microsoft has renamed Session Directory to TS Session Broker. (See http://technet.microsoft.com/en-us/library/cc772418%28WS.10%29.aspx)
To summarize, with Windows 2003 you have two choices when it comes to load balancing and session directory.
1) Have a HA implementation of Session Directory and use an external load balancer
2) Have non-HA implementation of session directory and use NLB to do load balancing.
Good how to article on setting NLB on 2003 with terminal services here:
http://www.brianmadden.com/blogs/brianmadden/archive/2004/11/29/how-to-configure-windows-network-load-balancing-for-pure-terminal-server-environments.aspx