Hi 6c1452,
Think you're in for the discovery of new boundaries in terms of BOINC. Not heard of separate boxes to appear as a single, multi CPU machine, but given that people build multiboard, multi CPU machines, I'd think it to be not impossible. Not sure about things like the uniformity of CPU's needed, the loss in traffic etc. Hope for you someone reports to the board to give an qualified opinion of the 'do ability' and practicality. Linux is free so you can configure it and clone it across multiple machines to your hearts content and all name them the same to make it appear they are one at BOINCstats, but think the servers will have a housekeeping problem if it all comes from the same member. Have you oogled for such a solution and plausability of implementing?

Just got back from the BOINC conference. There are a number of folks out there that are actually using their large clusters to run work from a BOINC project. They are generally submitting the BOINC core client as a job to their cluster. The core client starts up when it gets assigned to a computer and connects to the server, downloads work and starts running it. There are some problems with this approach that they are working with David Anderson to resolve (such as jobs submitted to a cluster this way just run until the kill time for the job is reached and then the workunits in progress are lost).

Another way that some people are using clusters to run BOINC is modifying the behavior of the grid software on the cluster so that it can manage whether the BOINC core client is running work or not (i.e. by suspending,resuming or somethings killing and restarting the boinc client).

Out of curiosity - what advantage are you looking for in terms of running those four machines as a one logical resource instead of just installing BOINC on each machine and letting it run?

It is a bit of a seeing-if-it-can-be-done project. The only real advantages to be gained from trying to use a cluster as a single logical device are ease of use and that the BOINC client's CPU number detection would work.

The BOINC client does seem to be multithreaded. Is it possible (out of curiosity) to force it to use a certain number of threads?


I think I may have just found what I'm looking for in Kerrighed:

"What is Kerrighed ?

Kerrighed is a Single System Image operating system for clusters. Kerrighed offers the view of a unique SMP machine on top of a cluster of standard PCs. "

...if that means what I think it means, and if it actually works. It does seem to be in development (and all the FAQs are for the 1.x versions and my browser can't render the information frame of the main page right now).

It's very close. Trouble is, though, apparently programs have to be written specifically for HPC in order to take advantage of the current clustering software.

Kerrighed seems to be the exception to this, but apparently it doesn't do SMP yet. I think I am leaning toward a multi-client farm approach until it does.