Use of diskless systems as number-crunching drone machines

[fatphil]

I have several machines whose only storage peripheral is the CD from which
they boot Knoppix. No floppy, no HD, no USB key. I use them as drones for
my number-crunching hobbies. I know space is tight, and I've read the FAQ,
but I'd like to suggest the addition of some number-crunching programs on
the live CD, so that I, and the few like me, don't always have to pull executables
off a central NFS server. In particular, the general purpose numeric package
'gp' (which is like 'dc', 'bc', or 'calc' on steroids).

Incidentally, I'm co-author of an upcoming (i.e. accepted for publication) article
in linux magazine where I explicitly mention Knoppix as my live CD of choice.

Keep up the good work, all those involved.
Phil

[A. Jorge Garcia]

What kind of number crunching are you up to? I'm playing with BCCD, clusterKNOPPIX, and QUANTIAN using openMOSIX and g++ to write clustered apps.

Regards,
AJG

[fatphil]

What kind of number crunching are you up to? I'm playing with BCCD, clusterKNOPPIX, and QUANTIAN using openMOSIX and g++ to write clustered apps.

I work on "embarassingly parallelisable" problems - prime number hunting. (Small packet problem description, lots of crunching, small packet of response - Amdahl coefficient of 1.0). (Google for 'fatphil PIES' finds my current big project.)

Presently it's a case of giving a job to a single machine, and then letting it run, but as my own client code is in active development I'd love to learn how to make its tasks farmable like true clustered apps. However, I don't want to oblige my clients to have to upgrade their clients just to run my app -- is there a cluster technology which is very unintrusive?

Phil

[A. Jorge Garcia]

Ah, very interesting! My students and I set up a 25 node openMOSIX cluster (each PC = PIV 2.6GHz + 750MB RAM, so cluster = 65 GHz and nearly 20GB RAM) and tried to figure out some number crunching program we could write that would stress the cluster. These students just finished a year of AP Computer Science using java. Unfortunately, JVMs will not migrate over MOSIX. So we looked into PVM and MPI but found the FORTRAN or C learning curve a bit rough for a 2-3 week final project.

So we wrote some C++ programs and see that they migrate well.

First we wrote the standard loop within loop within loop addition program to see if g++ compiled code will migrate.

We were looking into a monte carlo estimate of pi, but find writing a Very Long Rationals class a bit complicated.

So, we then we developed a LargeInt class, just for addition, to find very large Fibinacci sequence numbers.

Now, we too are looking for large primes. We followed the lead of GIMPS and are looking for our own Large Mersenne Primes. We extended that class from just adding Large Integers to also multiplying them and taking powers of the forms int*int=large, large*int=large, large*large=large, int^int=large and large^int=large.

We a little bogged down right now debugging that Large Integer class. Soon, we're hoping to conduct time trials to see how much the cluster helps to run this code as opposed to running on a single PC. What's nice about g++ on MOSIX is you can write and test programs off the cluster and then run them unaltered on the cluster. Maybe we could win a prize for the first billion digit mersenne?

I don't know what we'll try next. Any other similar prizes out there? My kids are very excited about all this!

Regards,
AJG