Path: utzoo!utgpu!news-server.csri.toronto.edu!rpi!zaphod.mps.ohio-state.edu!mips!apple!netcom!amdcad!dgcad!dg-rtp!matrx!wgate!lanzo
From: lanzo@wgate.UUCP (Mark Lanzo)
Newsgroups: comp.sys.hp
Subject: Re: CLUSTER MYTHS
Message-ID: <175@atesysv.UUCP>
Date: 18 Mar 91 14:25:27 GMT
References: <1991Mar14.231124.13611@ux1.cso.uiuc.edu>
Reply-To: lanzo@atesysv.UUCP (Mark Lanzo)
Organization: Wandel & Goltermann Technologies, Inc.
Lines: 56

In a prior article kemp@uiatma.atmos.uiuc.edu (John Kemp) wrote:
    
    I am trying to bunk or debunk a couple of myths about 
    HP clusters.  

Well, here's a couple of answers that should be good bunk :-)
Seriously, a little more info is really needed here.  What type
of cluster?  What revision OS?  The answers here are based on
HP9000/S-300 clusters with HP-UX 7.0B.

Disclaimer:  answers are based on my own experience;  any advice or
             information to the contrary would be greatly appreciated.

On with the show....

    1) Can devices on a cluster client be accessed by the 
       cluster server? (If not, hanging peripherals off of
       clients would be fairly exclusive.)

No -- unless you consider remote-login to the root server as "accessed by".
You can hang peripherals off of other nodes (for instance, our printer
is on one of our diskless nodes), but you cannot mount filesystems
locally.  

Access to devices on remote nodes seems to only work by having
a process actually running on the remote node which communicates with
your local process.  Of course, in some sense this is always true 
when sharing devices with foreign systems, what I mean here is that
this sharing is not transparent or automatic.  I am not referring to
file or device sharing protocols (like NFS, RFS, etc).  What I mean is
that programs like "uucico" turn out to be cnode specific files which
are actually stub programs -- these stubs run a server specific version 
of the program on the server and set up communications with it.

Sometimes this is a big problem.  For instance, we wanted to set up UUCP
connections over our phone line, and also needed a printer.  HP doesn't
support this -- both the UUCP and printer software want to run on the
server -- and there's only one serial port.  Buying an extra serial port
card for the server is not what I would regard as an acceptable solution,
given that we had plenty of free serial ports available from all of the
diskless nodes.  Fortunately in this case it wasn't very difficult to
con the system into running the "lp" scheduler on one of the diskless nodes.
On the other hand, we never did succeed in getting UUCP to work from a
cnode.
    
    2) Is a second network card really necessary on the server?
       (The manuals say "yes".  Other people say "no".)

No.  We have a bunch of workstations here, divided into a couple of
clusters.  None has more than one ethernet port.
    
    If there is a better source of documentation about 
    HP clusters somewhere, it would be nice to know about.  
    The Sys Admin manuals tend to be a bit weak in this area.
    
Yup.  The only thing I know of is the "HP-UX Clusters" chapter
in the "HP-UX System Administration Concepts" book.