[Beowulf] Beowulf - Warewulf ??

[Robert G. Brown]

On Wed, 14 Dec 2005, Rodrigo López Negrete de la Fuente wrote:

> Hello everybody,
>
> My name is Rodrigo, and I'm new to clustering so I've been reading about it
> and the available tools. The thing that's been troubling me is, that I
> haven't found what the main differences between Warewulf and Beowulf are. So
> is there anybody that can explain it to me, or is there a page where this is
> explained?

This precise difference isn't explained because it isn't really
something to compare.

A "beowulf" is technically a compute cluster engineered to be "a
supercomputer" -- single headed, with nodes that are the equivalent of
"CPUs" in an SMP system, using a commodity network as an IPC
interconnect.  These days "commodity networks" exist that are pretty
much designed for use ONLY in compute clusters of this sort, but that's
OK -- they (mostly) aren't terribly interoperable, but they are more
than one of them and they're competitive and they fit into "generic"
systems, so they are COTS parts (from a rather esoteric "shelf").

A proper beowulf cluster will USUALLY run a single problem at a time,
and in its most advanced version(s) -- Scyld -- the nodes run only a
stub operating system and you can't "log in" to them as if they are
computers any more.

Note well that the term "beowulf" long ago became sufficiently generic
that MANY clusters that don't match this description are called
"beowulfs", just as not all facial tissue is kleenex but nevertheless
get passed when somebody asks you to pass the kleenex.  So be prepared
for BOTH precise usage on this list AND for people calling a NOW, a COW,
a Grid, and all sorts of other clusterish conglomerates a "beowulf".

Warewulf, on the other hand, is a cluster computing software system,
specifically one that helps a novice user set up a fairly generic
cluster of systems that may or may not have disks as a diskless (PXE)
booted compute cluster.  The cluster booted may or may not be a
beowulf-type cluster -- warewulf would actually work fine to boot a Grid
style cluster, a NOW style cluster, or just a plain old diskless LAN --
depending on what images it serves to the nodes.  It is distribution
mostly-agnostic (by which it means that it tries to be agnostic as a
design decision, but leaves the problem of implementing on exotic ones
at least partly up to you).  It is not (generally) Scyld-like in its
sparseness -- you can log in to nodes booted from warewulf, depending on
how you set it up.

Scyld is "like" Warewulf but a) commercial, it costs money (although how
much can be dickered out based on a variety of factors); and b) it is as
described above a "beowulf" package, not really suited for a grid or
NOW.  It also IS a distribution in its own right.

There are several other "cluster support" packages and/or distros out
there, most of which support operations from "beowulf-like" through
"grid-like" on the whim of the configurer.

Finally, there is the mother of all unix cluster designs -- a bunch of
LAN workstations with remote shell capabilities, any sort of TCP/IP
network, and perhaps PVM or MPI installed.  Nearly all linux
distributions support this level and kind of grid/clustering right out
of the metaphorical box (or repository).  Hence my suggestion just a day
ago to another user to build a starter cluster by collecting 2+ linux
nodes and smashing a bottle of beer on one of them while declaiming "this
is now a cluster".

Oh, and using e.g. yum or apt to pop PVM and MPI onto them, as well...

This is enough to "get you started", although as you gain experience you
may WELL want to give warewulf, scyld, mosix, rocks a try.  Some
problems will give you incentive to do so.  Others won't -- they'll run
just fine on that network of workstations (NOW) cluster which is already
pretty simple to install and manage with most distros.

     rgb

>
> Any help will be very much appreciated!!
> Thanks
> --
> Rodrigo López Negrete
> http://muon.blogdns.org/~rush/
>

-- 
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