[Beowulf] about managment

[Robert G. Brown]

On Sat, 16 Oct 2004 llwaeva at 21cn.com wrote:

> Hi all, 
>   I am running 8-node LAM/MPI parallel computing system. I found it's
> trouble to maintain the user accounts and software distribution on all
> the nodes. For example, whenever I install a new software , I have to
> repeat the job 8 times! The most annoying thing is that the
> configuration or managment of the user accounts over the network is a
> heavy job. Someone suggests that I should utilize NFS and NIS. However,
> in my case, it's difficult to have an additional computer as a server.
> Would anyone please share your experience in maintaining the beowulf
> cluster? 

You don't need an additional computer as a server to run NIS -- just use
your head node as an NIS server.  Ditto crossmounting disk space with
NFS -- just export a partition (it doesn't have to be huge, just big
enough to hold your user home directories and workspace) to all the
nodes.  Remember, NIS and NFS were in some twenty years ago -- when I
first started managing Unix systems a "server" supporting dozens of user
accounts might have one or two hundred MEGABYTES of disk in exported
directories (an amount of disk that costs many thousands of dollars) and
might deliver a whopping 4 MIPS of performance, and yet the server would
still be useable for NIS, NFS and even some modest amounts of
computation in its relatively miniscule memory.

For a mini-cluster with only eight nodes, serving NIS and NFS won't even
warm it up, and the smallest disks being sold today are some 60 GB in
size -- an amount that even five or six years ago would have constituted
a departmental server's collective store (and that server might have
served 50 or 100 workstations, 100 or so users, and managed it with a
processor ten to twenty times slower).

There are also numerous alternative solutions, if setting up servers is
something you don't know how to do or don't want to do for other
reasons.  You could use rsync to synchronize user accounts across nodes.
This works well (and even yields a performance advantage) if they change
slowly, but will be a pain if they change a lot (the advantage of NIS is
that it pretty much completely automates this after a bit of work
setting it up originally).  You can and should use tools like ones that
were just discussed, e.g. kickstart and yum, to automate installation
and maintenance.  Finally, look into the various "cluster distributions"
that do it all for you, noteably ROCKS and warewulf.  Be aware, though,
that they are pretty likely to use things like NFS and NIS as (possibly
optional) components of their solutions.

[BTW, y'all are just spiled, spiled rotten.  Why, back in the OLD days
geeks were geeks and had to slam massive amounts of cola to get real
work done on networks, CPUs, memory that my PDA has beat hands down
today.  Here you put together a networked cluster the least component of
which would have been "inconceivably" powerful (in every dimension)
three decades ago, which is more powerful all by itself than the first
twenty or so beowulfs ever built, and you can't find a "server"...;-)]

    rgb

Robert G. Brown	                       http://www.phy.duke.edu/~rgb/
Duke University Dept. of Physics, Box 90305
Durham, N.C. 27708-0305
Phone: 1-919-660-2567  Fax: 919-660-2525     email:rgb at phy.duke.edu