[SGVLUG] linux cluster redo

[Michael Proctor-Smith]

On 8/2/06, jmd <jmd8800 at charter.net> wrote:
>
> On Aug 2, 2006, at 5:25 PM, Erik Espinoza wrote:
>
> > Hey Jeff,
> >
> > Clusters are a generic term for a type of technology. It really
> > depends on what you are trying to do. An Apache web cluster is
> > different than a MPICH farm or a parallel job farm.
>
> this was mentioned in the posts in the spring. i know very little
> about sys administration. this is a two fold project. one to actually
> sit down and decide what hardware suits my needs and then learn a bit
> about administering the cluster. and as said before, this is just a
> *home project*. my thoughts are to focus on doing
> climateprediction.net  for starters and use something like npaci-
> rocks as it come with everything needed like a good MPI.
>
> what i am looking at hardware wise is older opteron/xeon based
> systems with gigabit ethernet. (i certainly cannot afford an sgi
> altix with itaniums and NUMAflex interconnects.) one of the questions
> i kick around is, should i have a single processor motherboard or
> dual or even quad? i know in the sgi origins the NUMAflex
> architecture affords global addressing. without global addressing
> there might be serious latency issues. if i am using gigabit ethernet
> will i be having latency problems if i use 4 systems each with dual
> or quad cpus? it would seem to me that the cpu interconnects on a
> multiprocessor board would be much faster than the gigabit ethernet
> from board to board. so one board contains results and another board
> is still working. board one might be sitting there tapping its foot
> waiting for board two.
>
> getting back to the original question... i don't know enough to know
> what i want. :) something under 1000$ i can play with to learn. not
> being a computer professional i am not sure of what questions to ask.
> or proper terminology.

Your best is to figure out your application and ask the application
specific list what is the best hardware configuration. But ingeneral 1
socket baords are about half as expenise as 2 socket boards which can
be many times less expensise lthen 4 socket boards. So if your
application is not optimized for a cluster of smp nodes go with 1 or 2
cpu(exept that dual core and hyperthreading give you more logical cpus
per cpu socket). Just so you know in-general you have a problem first
then a solution that requires a cluster then you find the best
hardware for the solution.

> jeff
>
>
> >
> > Then there are specific app clusters, like some gentoo users use
> > distcc to compile from stage 1 in under 3 months.
>
>
> >
> > Erik
> >
> > On 8/2/06, jmd <jmd8800 at charter.net> wrote:
> >> last spring i asked a question about having a very fast head node in
> >> comparison to the other nodes.  that particular set of hardware was
> >> not working out (cheap stuff from ebay) and i canned the project
> >> until i got moved into a new place. since i parted those ibm eseries
> >> out to recover my money i need to look at buying hardware again.
> >>
> >> someone on this list mentioned they had a lot of experience with
> >> clusters. if i could, i'd like to pick your brain a little bit before
> >> i start buying again. if you are still interested in offering some
> >> advice please contact me off list.
> >>
> >> i know :)....i should archive my email so this request would not be
> >> necessary.
> >>
> >> thanks
> >> jeff duncan
> >>
>