Conference decwet::winnt-clusters

813.0. "Token Ring,CFS,NetBIOS,WINS,Clients,TimeOut,TCP" by BARNA::DSMAIL () Fri May 16 1997 08:02

  -------------- Long text follows. I hope it's useful for someone.
  ----------------- I hope it's not treated in other Notes entries.
  ---------------------------- Please excuse me for my bad English.
  ------------------------------- Don't answer to BARNA::DSMAIL !!!  

  Hello all...

  I've been suffering a special NT-Cluster installation and would like
to share this 'experience' with you all, and asking some questions (*>).
It's a bit long, but I expect you'll find it interesting.


  ----------------------------------------------------------------
  The client has three 'Token Ring' Rings (!). Only TCP/IP is used,
and each ring is a different subnet. 

  IBM has recommended him to put the two PCs which will be used for the
cluster IN THE THREE RINGS. It seems they want to avoid traffic across the
routers, but then, what are the routers used for ?
 
  So, each Prioris has 3 Madge Token Ring Adapters, each one in a different
subnet/ring. We managed to install the three adapters (one IRQ for each...),
and the private Ethernet adapter for the Cluster traffic. The two PCs have
NT 4.0 + SP2.

  The client swears that this configuration works OK. Apparently, everything
works, shares are seen, pings, traceroute and the like. They will change this
in one year for a Switch, but for the moment we must keep this configuration.

    (1>) (My first question : Is this configuration Useful/Usual/Recommended ?)
        
  -------------------------------------------------------------------

  Now, we install NT Clusters for NT v.1.1 in these two machines. The 
installation works OK, the shared disks (RaidArray 310) are seen, the failover
groups are created, Failover/Failback works. However, from this point on,
the USer Manager for Domains becomes VERY SLOW. (One machine is PDC, other
BDC. Getting the User List in PDC takes 45'')

  After reading Note 345, and many tests, we disable CFS.SYS. The Cluster
keeps working, and now performance is back to normal (1'', User Manager).

    (2>) What is CFS.SYS used for ? 
        (Apparently, it's only used to slow down User Manager :-)
    (and 2.bis) Where can we get REALLY TECHNICAL INFORMATION about these 
   drivers, client part, NetBIOS names used, and so on ? (The PPT presentations
   and the manual are not very technical, indeed)

  -------------------------------------------------------------------

  As clients, we use one Win95 + SP1 machine, and one Windows For Workgroups
(+ Network driver patches from the NT 3.5.1 CDROM), both of them in one ring.
We install the cluster client software on both.

  They connect OK to the cluster shares. We do a Manual Failover, and the clients
get more-or-less hung. The Win95 client can sometimes reconnect, the W4WG one
usually hungs. If we delete the NETBIOS CACHE (nbtstat -R), the reconnection
can be made without any problem. If not, it usually hangs.

  NET LOGON from the clients (from the graphic interface) is terribly long.

  We do two changes then :

      1. Changing the Client NETBIOS Resolution mode to b-node (Mode 1 via DHCP,
  could be changed via SYSTEM.INi or Registry, see Stars or Technet), and,

      2. Changing the two Timeouts in the client (SYSTEM.INI, Registry, 
  according to the NT Cluster manual, you must create both entries, by default
  they don't exist), with values 1 and 1 (which we thought was the minimum),

  THE CONNECTIONS NOW WORK OK IN BOTH CASES. Failover is almost immediate.
 File Manager rebuilds the directory list for the Cluster shares immediately.
 Net Logon is immediate, too. (There are only 4 PCs on the ring)

 The customer begins to be happy. But we have some more questions :

   (3>) We'd like to use WINS, not only b-node.

       Is it supported ? Does it work ?
       How do the clients and the server use the Cluster Alias ?

          If it's a normal NetBIOS name, it should be immediately refreshed
       on the WINS Server, which it seems not to work.
          Should the Cluster Member be WINS Clients to achieve this ?

          In fact, we'd like to know how NetBIOS uses names related to the
   cluster and how and when the cluster changes them (the IP address associated)


   (3.bis>) Does the Cluster send a R-ARP to notify the clients ?
           Is there something similar for NBTSTAT ?

   (3.bis.bis>) What do the two timeout values mean ?
                Why do they affect to other client components (NET LOGON) ?
                How to configure them on Token Ring (in Ethernet, there
            was no need to change them)


   ---------------------------------------------------------------------
   We don't see any need for Cluster Client Software. Depending on how
 NetBIOS names are created/used, the Cluster servers should do all the work.
 We think that we'll be implemented in NT 4.0 SP3 + Cluster 1.1 SP1.
 
   (4>) Is that right ?
        Any technical info about this change / new philosophy ?
        Can we get a Beta already ? 
        When will it be released ?   


  --------------------------------------------------------------------------
  Many thanks in advance, and greetings from Spain

  Juan Carlos Ruiz
  ruizju@mail.dec.com
  Digital MCS, Barcelona, Spain


  
  

  

>  The client swears that this configuration works OK. Apparently, everything
>works, shares are seen, pings, traceroute and the like.
    
    That is great!
    
>(1>) (My first question : Is this configuration Useful/Usual/Recommended ?)
    
    It is useful for your customer.  It is not usual in our experience.  We
    have not tested any token ring adapters for cluster use, but they
    should work ok.
    
>(2>) What is CFS.SYS used for ? 
>        (Apparently, it's only used to slow down User Manager :-)
    
    Yes.  That's why we wrote it :-).  "CFS.SYS" is the "cluster name
    client" redirectory for NT.  It allows an NT system to connect to a
    cluster using a "cluster" name instead of an individual node name. 
    Another piece of cluster server software, "CNS.EXE" runs as a service
    on NT Cluster servers, and interacts with CFS.SYS on client systems to
    get you connected to the correct server in the cluster.
    
>   We don't see any need for Cluster Client Software. Depending on how
> NetBIOS names are created/used, the Cluster servers should do all the work.
> We think that we'll be implemented in NT 4.0 SP3 + Cluster 1.1 SP1.
> 
>   (4>) Is that right ?
>        Any technical info about this change / new philosophy ?
    
    I skipped lots of your questions to answer the most important one. 
    With NT 4.0 SP3 and Cluster V1.1 SP3, you don't normally need to use
    the cluster client software if all your servers and clients use TCP/IP
    and NetBT (Netbios over TCP/IP).  You must configure "IP Failover"
    objects in your cluster, and put an IP failover object into each
    failover group in the cluster.  Clients should connect to the IP
    address and Netbios name contained in this IP failover object instead
    of using individual node names.
    
    Clusters V1.1 sends R-ARP messages when a "floating" IP address moves
    from one node to the other.  V1.1 SP1 adds support for Netbios over
    TCP/IP (NetBT), which can be used for file share and named pipe
    connections to the cluster.  It notifies WINS servers of the Netbios
    names that are created.
    
    Previous cluster versions used the cluster name server/name client
    protocol, which only works on LANs, and does not use WINS or ARP at
    all.  We still provide this software, but we recommend you use the new
    TCP/IP functionality if you can.  As you noticed, the cluster name
    client software can slow down network connections.