Beowulf vs Extreme Linux

[Pete Beckman]

At 10:40 PM 9/1/98 +1000, Jacek Radajewski wrote:
>What is the difference between Beowulf and
>Extreme Linux.

The Beowulf community has centered on exploring commodity hardware and
software for parallel processing, focusing on "How to Build a Beowulf" and
the software, networking, and technology required to run parallel programs
on that type of cluster.  The Beowulf community is best known for building
parallel computers from individual parts (e.g. mail-order supercomputer),
and the networking, drivers and software (ethernet, channel bonding,
distributed shared memory, PID space, etc), produced by CESDIS, and Donald
Becker et. al.  Important design goals seem to be "commodity hardware" and
"free software".

It is my understanding that a Beowulf can be constructed from ANY commodity
hardware and freely available Unix, not just Linux.  A cluster of
Pentium-class machines running NetBSD, FreeBSD, or Hurd could certainly be
a "Beowulf".  Sun recently announced solaris is free for non-commercial
use.  I'm guessing that stringing together a bunch of commodity Solaris-x86
boxes with PVM/MPI would be a Beowulf.  On the other hand, a cluster of HPs
running HPUX would not be a "Beowulf", since the hardware nor the software
is commodity.  However, surely debates about what is and is not a Beowulf
are pointless, and waste everyone's time.  The point is that the Beowulf
community focuses on inexpensive commodity hardware and software to
construct parallel computers.

People discussing Extreme Linux focus solely on Linux, and pushing it to
the edge.  Commodity, "mass-market", or free software, are not central
themes.  Many people focus on high-performance scientific computing, but
the goals of high-performance computing often overlap with many other
areas.  Discussion covers desktop linux compilers and tools, non-commodity
networking software, real-time linux and instrumentation, models of
parallelism, and parallel graphics, etc.  The availability and utility of
commercial compilers, tools, and other software is discussed.  Buying
complete 'ready-to-power-on' Linux clusters is also a reoccuring theme, and
often encouraged.  The Extreme Linux community does not have a particular
client/master-node view of parallel computing, and does not focus on "How
to Build a Beowulf", but rather Linux.  Folks with an interest in parallel
scientific computing (I'm in that group) ask questions like:

 - How can the Linux kernel be trimmed down to support very thin compute
nodes?
 - How can multiple graphic cards be joined to do parallel rendering?
 - When will MetroLink release a Linux driver for ultra-high-end graphic
cards ($3000+)?
 - What does the kernel need to support efficient put/get over DMA-capable
network adapters?
 - Who has HiPPI adapters working in Linux, and connected to HPSS?
 - How do the KAI and Portland Group compilers compare?  How about the
development environment?
 - Are the compilers availale for Alpha Linux?  
 - When will DEC make it's scientific math libraries for the Alpha
available to Linux?
 - What does it take to make Linux a supported platform for scientific
computing at a Laboratory?
 - How can the kernel instrument threads on an Linux SMP box?
 - Is Linux support stable enough to gain acceptance from management as a
contender for a procurement bid?
 - How well does LSF's commercial product for job queuing work on a cluster.


etc...

-Pete


---
========================================================================
| Peter H. Beckman                   |  Advanced Computing Laboratory  |
|  Los Alamos National Laboratory    |     Phone: 505-665-0800         |
|  CIC/ACL MS-B287                   |     Fax:   505-665-4939         |
|  Los Alamos, NM 87545              |     email: beckman@acl.lanl.gov |
========================================================================