A small little survey

[Dan Kirkpatrick]

>Date: Thu, 21 Jun 2001 10:20:02 -0500
>From: "Jared Hodge" <jared_hodge at iat.utexas.edu>
>To: beowulf at beowulf.org
>Subject: A small little survey
>
><snip>
>The questions:
>
>What is your hardware setup (processor, motherboard, network, etc.)?

16 dual Pentium 300Mhz, Supermicro P6DLE, 100Mbps switched, all IDE except 
head node->, scsi.

>Did you have any problems with the hardware or configuration?  What is
>your evaluation of the cause and solution?

Bad experience w/ Supermicro.  have had 3 out of 16 boards fry within 2 
years.  1 within warranty and they couldn't even replace it.  Also, the 
cheapo place that built 'em (mini tower PC vendor) put cheap heat sink/fans 
on processors... most stopped working within 1.5 years.  Replaced all of 
them with fans from co-fan usa (http://www.cofan-usa.com/) with MTBF of 
50,000hrs instead of the usual 10,000 hrs.
Hardware monitoring will be a must for next cluster.  Good thing we had 
plenty of cooling in the room since never had a processor burn up with the 
fans going dead.

>Are you pleased with the performance and reliability of your cluster?
>Please explain.

Performance for the cost... excellent.  Reliability was "fair"... Our 2nd 
attempt now with 4x the budget will not go with "cheapo" boxes... going 
reasonable rackmount systems from clustering vendors.

>Assuming you have a limited budget (that should be everyone), what
>recommendations would you have concerning cluster hardware?

Typical setup we're looking at:
head node scsi, slave nodes IDE, all 100baseT switched, dual Pentium or 
Athlon, rackmounted (not much more than minitowers)
no kvm switch, we can move KVM cables manually... no fancy 1U LCD 
console... full monitor/kbd/mouse, no UPS, no expensive clustering software 
solution.

>What has been your experience with the following hardware and financial
>tradeoffs (assuming you've had some experience with them)?
>Intel vs. AMD processors:
>SCSI vs. EIDE disks:
>Myrinet vs. Ethernet only vs. other network technologies:
>Server level hardware (such as the supermicro 370DLE motherboard) vs
>workstation level hardware:

Intel vs. AMD... not quite sure yet, hear you get more bang for buck with 
AMD, but to really take advantage... go dual processors, but AMD's are not 
quite proven, sometimes not available, yet.  We may go the AMD route on 
cluster 2.
100baseT Ethernet best cost if not parallel computing.  Probably bigger 
coast savings here.
Scsi may add $200 per node... needed if disk is the bottleneck.  If using 
head node with NFS disk, then net is prob the bottleneck.

You dont mention what the main need is for... threaded or parallel 
computing... memory or disk intensive?
These really drive the best bang for the buck.


=======================================================
Dan Kirkpatrick                   dkirk at physics.syr.edu
Computer Systems Manager
Department of Physics
Syracuse University, Syracuse, NY
http://www.physics.syr.edu/help/    Fax:(315) 443-9103
=======================================================