Googled this a little while ago, so googled it again: http://boinc.berkeley.edu/dev/forum_thread.php?id=8172#47568
http://forums.anandtech.com/showthread.php?t=81515
http://comments.gmane.org/gmane.comp.distributed.boinc.devel/4997 (with comment by BOINC developer)
and more at https://www.google.it/search?q=boinc+on+beowulf+clusted

Bon chance with your challenge (No idea if this works at WCG, but it will have to identify with a proper .gnu Linux platform to get work)

The Pineview Atoms are interesting CPU's and I have built passive cooled servers with them using Intel MUD motherboards.

However, in terms of work done per Kilowatt/hour of power used, and the capital cost of purchasing Atom motherboards and PSU's, the lower power dual core Celeron Socket 1155 CPUS offer around four/five times the (raw) benchmark capacity, or better. (I have not tested these with BOINC) The LittleFe Atom cluster's prime requirement was portability for educational use.

The price of a Celeron G1610 or similar and a 1155 motherboard is around 30% more expensive than Atom CPU equipped motherboards at retail rates, but there are bargains everywhere. Given this, a cluster using half as many Socket 1155 MATX mobos and Sandybridge Celerons might get a lot more work done than using Pineview Atom CPU's for the same, or slightly cheaper cost, and the motherboards can be upgraded to Sandybridge CPUS like the i3/i5/i7 as the prices of these CPU's fall drastically next year. The draw of the Atom chips is their low power consumption, Mini-ITX motherboards (and passive cooling) but this does equate to (comparatively) less crunching ability. Also memory cost and performance in clusters is an issue, and even Atom powered machines usually need fan cooled PSU's

The use of Atoms has been discussed hereabouts if you search; essentially, for low powered single CPU systems the trick is to justify the lesser number of work units crunched for the power used and capital expenditure, while at the same time making sure that work units are completed before the deadlines expire. This latter requirement ruled out my use of Atoms for the experimental BOINC crunchers built last year.

Another view, given that an i7 quad core cpu can chew through eight tasks at once and produce remarkable results per hour of runtime, would be to compare the output of just one of these with say, a cluster of four Atom dual-core boards. The power consumption of an i7-3700K at 100% BOINC over eight cores is just 160 watts at the wall socket.

After that comes the use of graphic card processors; while there are no current work units available for GPU there could be in the future and some ability for use of GPU's to be added might be worthwhile.

Of course, some-one now has to benchmark these comparisons (you?) even theoretically, and the results will certainly be interesting. My own research (April 2012) led me to believe that an i7 running eight tasks at 100%, 24/7 would be more productive than 90% of the (single cpu ) WCG machines running at that time. The only issues are case design and cooling considerations. Whether that single machine would be faster than an Atom cluster running, say eight dual core CPU's is interesting, to say the least, but my money is on the i7-3600K.

Of course, building 'another desktop PC' is less interesting than nailing together a commodity cluster, but for the same 160 watts of electricity, you're going to have to come up with some interesting science to get close to the results from an i7 or better using clustered Atoms. If power consumption is not an issue and it usually becomes one, then a cluster - or rather the results from it will be very interesting. However the Atom series of CPU's can be significantly bettered for very little in the way of increased expenditure if you're building a research machine/cluster.

Regards

You're welcome. There are some interesting threads in the Chat Room about multi-CPU (4 CPU) motherboards aiming for the top 5 fastest machines on WCG, but of course you may be heading for other uses than WCG.

In terms of clustering for WCG all of the task loading, scheduling and 'node control' are handled by IBM & BOINC, saving us all the trouble of building Beowulf clusters, but of course, the notion of a 40 watt per node super-computer cluster under the bed is an appealing one.

It is possible to build almost noiseless high power machines, water cooling is one option amongst many and there's plenty of advice on here.

I double checked my own Atom device, built for WCG but too slow to turn in results speedily,it's one of these , and now does duty as a Netflix client in the living room.

The later Atoms benchmark almost as fast as the Pentium 4 Northwoods we were all so pleased with in 2004 or so, and cannot be deprecated, but 2004 was a long time ago in CPU terms.

Also, while many of the later Atoms are 64 bit enabled, the BIOS on many of the passive cooled motherboards, and also inside netbooks will not allow 64 bit OSs to be installed, which is to say the least, inconvenient.

Having said that, my current laptop is a 2004 ThinkPad T40, so there is plenty of life left in older CPUs, until we take FLOPs/$ per Kilowatt hour into consideration in soon to be complete 64 bit world.

Hi Pirre eu,

Have you used/tried the Dotsch_UX type of server/client diskless cluster?

I know it is not Beowulf/HPC but it is another option.

Yes! I experimented with it a couple of years ago.. It is an old distribution of Ubuntu, but works fine as proof of concept.
You can set it up in a day and have a disk based Server control any number of diskless/headless Clients.
Its lots of fun and cheap to set up and play with, if you like the basic setup you could use a more modern distribution.

Have fun