[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Getting Started

Thanks for the reply Jorge,

My first impressions mirrored your points exactly, each of my cluster nodes has 256Mb of RAM (thankfully the donators were a financial firm who loaded all the machines with the RAM).

I am very envious of your cluster, each node sounds quite powerful, much more powerful than my severly dated machines. However, we must all start somewhere and hopefully this will appeal to the 'powers that be' and also be useful for the undergrads for initial steps into leaarning parallel programming concepts.

I am in the process of creating my node kernel as we speak, I am using the

make-kpkg --append_to_version diskless buildpackage

I have been doing alot of digging, and I am forcing myself  to do things the right way the first time. From what I have managed to dig up, the make-kpkg tool is the best for a debian system. I know I will not be working on this machine forever and I want to make a new admin (hopefully an undergrad student) as comfortable as possible, with as much documentation as possible.

I assume you are using debian, and did you use this make-kpkg process? I know it's a one time thing, and once set up is not neccessary to modify, but if you recall what was done, it would be helpful to me.

Cheers,
Dan Nedelko
<a href="" class="moz-txt-link-rfc2396E" href="http://dannynedelko.homeunix.org/modules.php?name=Sections&sop=listarticles&secid=4">"http://dannynedelko.homeunix.org/modules.php?name=Sections&sop=listarticles&secid=4">My Beowulf Journal</a>



Jorge L. deLyra wrote:
2. I am limited to diskless nodes (due to $$$) however, what in your
experiences provide a better setup? disked machines which need syncing
or diskless nodes?

We have several diskless clusters around here working _very_ well. In my
opinion diskless nodes for a number-crunching cluster are _much_ better
than nodes with disks because:

- They are simpler and cheaper.
- They are _far_ more reliable, hard disks are trouble.
- They boot faster from a 100 Mbps network than from disk.
- Their system administration is easier, all done on the server.
- They can boot from a boot prom and have no magnetic media at all.
- If it comes to the worst, you can just reset them without problems.
- They don't use swap, which you don't want because, if your numerical job
  begins to paginate or swap, you are dead performance-wise. Don't even
  consider doing swap over the network. Memory is cheap these days, just
  put some 256 MB in each node and forget all about it. 

Here is the recipe for a typical node of ours:

1 clean motherboard (the less things onboard, the better)1 CPU (Athlon 1200 MHz with a good ball-bearing fan)
1 SDRAM DIMM (256 MB is quite comfortable)
1 network card (3C905C with the flash boot prom)
1 network patch cord
1 power supply (standard ATX)
1 power cord

This is it, we don't even use cabinets, just stash them on shelves. We
don't use any LEDs or switches, we configure the motherboards to come up
when powered on. If anything goes wrong we power-cycle the node unplugging
the power supply. It's a bare-bones setup, but it works _very_ well.

							Cheers,

----------------------------------------------------------------
        Jorge L. deLyra,  Associate Professor of Physics
            The University of Sao Paulo,  IFUSP-DFMA
       For more information: finger delyra@latt.if.usp.br
----------------------------------------------------------------




Reply to: