[Beowulf] Should I go for diskless or not?

Gus Correa gus at ldeo.columbia.edu
Thu May 21 20:02:11 PDT 2009

Dr Cool Santa wrote:
> A big discussion I see . . .
> Finally I was able to get the diskless set up running, and the machines 
> did come with disks . . . I set it up diskless just to save time.
> And actually Bio Chemistry work is going to be performed on this 
> cluster, but I don't think a lot many people would be working on it at 
> the same time, hence we aren't going to have a big cluster of nodes, 
> probably not even hundred for that matter. As some of you said already, 
> I could create local swap partitions on each but I wanted to make one 
> computer have all the hard disks so its not really hard to you know 
> access the left over memory. The computers come for the company, so the 
> hard disks have a decent amount of memory I believe (around 300GB each). 
> Since this is diskless at the moment, and the virtually each machines 
> hostname and ip address changes according to the order they are switched 
> on, you can think how hard it is going to be to access the hard disks 
> from some node. I already said, we aren't making a super computer, its 
> only going to be used by one research group, the university provides us 
> super computers to work on but people here just love to destroy other 
> people's research work.
> I want to emphasize that it is a small cluster and will remain quite 
> small, i can just shut the machines off by pressing the buttons without 
> even moving an inch, and of course we always have SSH to do the thing 
> for us.
> Some sophisticated things you guys mentioned there, but I doubt we'd 
> need it.
> And at the moment, there is no administration software installed for the 
> cluster, I am actually just working on about what the software wants me 
> to install. THats MPICH1 for one and HPMPI for another. I am new to 
> linux networking and to clusters, my ability comes from tutorials which 
> are scarce and I don't know how outdated (decades to be precise) or I am 
> looking for them in the wrong place. Therefore if you could give your 
> suggestions to this newbie, it'd be appriciated.
> And to mention, we are using Schrodinger suit of software and Accelry's 
> Materials Studio here.

Hi Dr. Cool Santa

I suggest that you try OpenMPI, MPICH2, or MVAPICH2, instead of MPICH1.
MPICH2 runs on tcp/ip (Ethernet), MVAPICH2 on Infiniband,
and OpenMPI runs on both (and on Myrinet).
Hence the choice depends a bit on the network hardware you have.

The MPICH developers recommend moving on to MPICH2.
MPICH1 is not maintained anymore, and it seems not to work with
current Linux kernels.
Kids and adults don't believe on it, not even in North Pole.

Get the latest OpenMPI, MPICH2, and/or MVAPICH2 tarballs
on their web sites:


Install MPICH2 with the nemesis communication channel,
which is not the default, but it is the one the developers recommend,
and also works best.
OpenMPI and MVAPICH2 are very easy to configure and build.

MPI is a standard API, hence your programs should compile with all of 
them, as long as you configure them right.

I hope this helps.
Gus Correa
Gustavo Correa
Lamont-Doherty Earth Observatory - Columbia University
Palisades, NY, 10964-8000 - USA

More information about the Beowulf mailing list