I remember when the event was about to happen and almost drove up to it, but recall having to work that weekend or being away or something like that.  This was nearly 3 years ago, in April of 2004.  The last update to the site was in May of 2004 and Google only (currently) has 17 pages for "flashmob computing." That's pretty sad, actually.

Their goal was to break in to the Top 500 supercomputer list.  Three years ago, this would have been possible with a supercomputer with less than a teraflop of performance (and to think Intel demonstrated a single CPU recently that could do that).  Last year, the mark was already up to over 2.7 teraflops.  That's just to make it to position 500, which wouldn't last very long at all.  Although more than 700 computer arrived, they never used more than 256 of them that were rolling at around 180 gigaflops.  The final number was with only 150 computers and 77 gigaflops.

That was three years ago.

I downloaded the ISO that they used.  Instead of burning it to disk and booting off of it, I wanted to try it within VMware as I have this notion that I might be able to use it on EC2 for a fun test.  It booted fine.  It only has flags for up to a P4 system and there are some autodetection issues with the Intel Core Duo I have.  Aside from that, it ran fine with a single CPU and gave me a 1.35 Gigaflop result inside a VM with only 256MB of RAM.  I brought up a second one, with the two on their own network.  This resulted in full use of both Cores and a results of about 2.35 Gigaflops.  This is on a single CPU, with 2 1.66Ghz Cores (first gen) running on a 12" laptop inside of two VMware VM sessions!  Three years ago, 2x 2GHz celerons and 2x 1.8Ghz P4 systems only produced a result of 2.35 Gigaflops. (I'm actually itching a bit to try this on our QX6700 system, but that'll have to wait a bit.)

It surprises me how few people sent in results since this is so easy to run and presumably even easier if you burn the CD.  Maybe 200MB was too much 3 years ago for most?

Without spending any time on it, I'm trying to make sense of what the boot CD is doing and how I can maybe convert it over to a more modern version of Linux.  It's running on a pretty stock Knoppix boot disk with kernel 2.4.  I don't think there is much to it other than a couple of directories.

It seems like technology like this would still be in active development.  Reading the difference between this and Grid computing makes it seem like this sort of solution is much more general purpose.

The closest thing, on the Grid side, might be MacResearch's OpenMacGrid, which is a distributed computing environment with a difference.  It allows any research to submit projects for acceptance on to the grid.  It's unclear, though, how something less popular than SETI@Home or Folding@Home could gather enough people to have a really fast and reliably fast distributed computing environment.  You also can't really have more than one such client running as they all typically run in idle cycles. 

Google actually used to have something similar. Although it wasn't openly accessible, it wasn't as narrowly focused as most of the distributed computing clients.  According to the Wikipedia entry about Google Toolbar, Google Compute ended near the end of 2005 and was never used for more than assisting with Folding@Home. 

The Wikipedia entry on "flash mob computing" has only duplicate information from the USF site.  It's as if the failure to break in to the Top 500 list killed the entire concept.  I've tried contact them.  If I hear anything, I'll add it to this article (and so far I haven't).

A SourceForge search pulled up Cbench which led to information on the Thunderbird Linux cluster at Sandia. This led to Open MPI, an open source project to implement a standard API for the Message Passing Interface as well as OpenFabrics, which gets even more beyond what I can understand in just a few minutes.  The Thunderbird Linux cluster currently runs at 58 Teraflops with 8694 processors across half as many machines. The pictures on their site show that it is clearly an exercise in high performance networking more than anything, as the FlashMob Computing folks learned at their trial.

Sure, given enough time and experience, one could probably develop some sort of boot disk that used these tools, but I still think it might be easier to modernize the FlashMob one. ;) Sandia, after all, has the funds to do as they wish.

Posted by Shane on February 22, 2007 8:17 PM | Permalink