[dsp-clusters] Question

Gerry Creager N5JXS gerry at cs.tamu.edu
Thu Aug 17 06:24:59 PDT 2000


For the cluster I've beenworking on at home (16 P133 nodes) I've gone
rack-mount.  

For the cluster I'm building at the office (about 60 nodes,
heterogeneous 486/33 thru PIII-667) I'm using racks with system
shelves.... but only because my boss decided that wire racks were not
nice-looking enough for the tours we lead thru the new network
engineering lab.

The home cluster is constructed of 4/chassis P-133 SBCs on an ISA bus
with 64MB of memory and 4GB of local IDE disk.  I got chassis at auction
for $75 each, and was the benefactor of the largess of a telco vendor
updating their P133s to PIIs last summer.  Mechanical KVMs were $25/4
ports at the Dallas, TX "First Saturday" giant fleamarket.  The rack was
another $75 for a scavenged AS-400 cabinet.

The home cluster was done this way primarily for space management in my
shop.  Aside from the 16-node cluster, I have several other boxes
external to the cluster for ham radio, meteorology and realtime
monitoring.  The thought of scattering 20-odd tower cases around in
there was not pretty.  And, I got everything cheap.  Note that I got
enough cases to put all the CPUs in the rack.  While they're on a
dedicated and isolated 20-Amp circuit, power is still a concern.  I'm
max'd out for that one practically, if not theoretically.

For the office, the issue is deployment of a number of tower-case
systems.  With the number I have to work with, I'd have preferred to
stack 'em on tables, or wire shelves.  More space-efficient.  But, to
maintain the aesthetics, we're using system shelves in open-frame relay
racks.  Also for aesthetics, my switches are not colocated with the
systems although they're in the same room.  They're in a seperate set of
racks dedicated to network devices.

Thus, the decision to chassis/rack mount or not is determined by a
number of factors.  For my work at home, the primary consideration was
space.  At the office, building a working cluster is secondary to the
need of having a new laboratory that is also a showcase for our network
engineering program.  Were I building a cluster solely for for computing
(soon, now, soon) I'll revisit the considerations, and suspect I'll fall
on Eugene's side of the equation.

I feel compelled to remind everyone that for the TCP/IP seven-layer
model, 2 new layers have been identified to be added at the top (thus
creating a 9-layer model...):  Layer 8 is Financial, and Layer 9 is
Political.  I believe that Layers 8 & 9 are present inthe decision of
how to configure a Cluster environment, as well.

Regards,
Gerry
--
Gerry Creager		gerry at cs.tamu.edu, gerry at page4.cs.tamu.edu
Network Engineering			|Research focusing on
Computer Science Department		|Satellite Geodesy and 
Texas A&M University			|Geodetic Control
979.458.4020  (Phone) -- 979.847.8578  (Fax)




More information about the Beowulf mailing list