A small little survey
RSchilling at affiliatedhealth.org
Thu Jun 21 09:48:49 PDT 2001
I've worked in a variety of clustering environments, Beowulf and otherwise.
Depending on what your budget is, you might look at the following platforms
(in order of proven
Compaq Alpha systems, and IBM
The advantage you have in working with these vendors is that the systems you
get are complete hardware wise. They can easily be compared to each other
in terms of performance because there is less variability in hardware
quality compared to building your own from scratch (e.g. separately
purchased parts). When figuring cost, don't underestimate the cost of
tinkering with hardware and dealing with various vendors at the same time to
purchase your equipment.
What is your hardware setup (processor, motherboard, network, etc.)?
We currently have a couple of clusters: One with recycled Compaq PCs
(Intel) of PII/PIII vintage, another with Data General PCs (Intel) with PIII
- currently being installed.
Did you have any problems with the hardware or configuration? What is
your evaluation of the cause and solution?
Not too many problems at all. Just plug them in and they work for
the most part. Biggest configuration issues have come out of configuring
the software to "see" the hardware properly - normal system admin stuff.
Are you pleased with the performance and reliability of your cluster?
Pleased for what we are using them for. The first cluster I
mentioned above is for file storage and experimental application
development. The second cluster I mentioned above is currently being
installed as our main information system, but other companies who have
installed the system are happy with the performance. The second cluster
will support an enterprise environment with 1,200 users, and various web
users just fine.
Assuming you have a limited budget (that should be everyone), what
recommendations would you have concerning cluster hardware?
Try no budget here. My primary source for hardware currently is the
supply of used workstations and networking equipment that are being replaced
in my company. They're typically only a few years old at best, and work
What has been your experience with the following hardware and financial
tradeoffs (assuming you've had some experience with them)?
Intel vs. AMD processors: Intel tends to be a little more expensive between
the two. But, you should also weigh the cost of hardware and people-hour
cost of purchasing AMD or Intel baseed nodes in pieces -vs- the cost of
Sun's new workstations. They have a basic 64-bit workstation out now for
$995 (called Sun Blade). As of this writing you can see it at
http://store.sun.com/catalog/doc/BrowsePage.jhtml?cid=60357. It runs 64 bit
UNIX and I'm certain it would be fine for clustering. You might also check
out the Mac G4 running MkLinux. Not sure if Mac OS X will lend itself to
SCSI vs. EIDE disks: I'm a big SCSI fan, but EIDE works fine, especially if
you go with a RAID setup. I think it's more of a question if wether or not
you'll have to repurchase new hard drives every time you move to a different
motherboard. Use whatever your motherboard vendor plans to use in the
future. It really nice to be able to move hard drives between nodes,
especially when you use recycled equipment.
Myrinet vs. Ethernet only vs. other network technologies: not sure.
Ethernet seems to work fine.
Server level hardware (such as the supermicro 370DLE motherboard) vs
workstation level hardware: On the PC side, I'm not sure you'll see much
difference, but others on this list would know better (especially Mr.
Brown). You will see a difference in things like throughput with the other
machines I mentioned (Sun, Alpha, IBM, etc . . .).
Hope that helps.
Webmaster / Web Integration Programmer
Affiliated Health Services
Mount Vernon, WA
From: Jared Hodge [mailto:jared_hodge at iat.utexas.edu]
Sent: Thursday, June 21, 2001 8:20 AM
To: beowulf at beowulf.org
Subject: A small little survey
Ok, I've got a bit of a problem and I need everyone's help (everyone who
is running a cluster at least). We are considering buying a cluster,
but we need to stick with proven hardware. We've been considering using
the supermicro 370DLE motherboard with 1 GHz PIII processors (1 per
board) and Myrinet 2000. Our cluster will probably consist of 16 nodes
and one server. I'm wanting to get responses from people who have
similar and/or very different configurations (AMD chips, ethernet
only). We will be using our cluster for finite element analysis, and
our current cluster consists of 8 PII 350s plus server and we are using
a 32 bit 33 Mhz PCI interface with the older Myrinet (1.2 Gbps max - not
really half that with 32 bit PCI). Personally I'm wanting to explore
the possibility of getting AMD chips but I'm told it's not a "proven
technology." I would like to get a mass of responses from people who
are running clusters for scientific research and find out what
technologies are proven. I'm personally not biased either way, but I
just want the most performance and reliability for the money. I know
that the following questions are very vague, but I need responses from
people who have used the hardware and can share their experience. Feel
free to add comments. Please answer any or all of the questions in a
reply to me (you can reply and remove beowulf at beowulf.org as a recipient
since I don't think everyone wants to hear all of the responses). I'll
summarize the responses in a mailing back to the list at a later time if
it is appropriate.
Thanks for your help.
Institute for Advanced Technology
The University of Texas at Austin
3925 W. Braker Lane, Suite 400
Austin, Texas 78759
Email: Jared_Hodge at iat.utexas.edu
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit
More information about the Beowulf