I ran across a site at Oak Ridge National Laboratories (http://stonesoup.esd.ornl.gov/) that describes a Beowulf-class (Linux) supercomputer running GRASS that the rocket scientists there had made for free out of junker PCs and open source compilers.There are a lot of problems in GIS especially suited to parallel computing.The idea of building my own supercomputer out of the junk I'm currently using for storeroom stuffing and free software I can download on the Internet has some real aesthetic appeal.
Curious to know if anyone else besides ORNL has tried to use recycled computers in any way to work on GIS problems, I asked the GISList (www.directionsmag.com/discussion) readership, "Problems in modeling, image processing, land use classification, remote sensing analysis, etc.etc.are classic examples of GIS problems that are well suited to parallel processing techniques.I would like to hear from anyone in GIS who is looking at these open source, recycled PC solutions."
This idea excited the kid in me who still believes in the Phoenix, the bird of myth who rises from the ashes of its ancestor.But cooler heads than mine dipped that idea in an ice-water bath.Well okay, only an incurable romantic whose head is on fire and whose feet are grounded in the clouds would try to build a supercomputer out of discarded junk.To those fools of science who ignore conventional wisdom, may your tribe increase! Sooner or later one of you will prove conventional wisdom wrong.
Academia is where the Phoenix lives (but there also lives Prometheus and the spark that ignites the light of science), so the first suggestion by Richard Hoskins makes good sense: "Beowulf stuff is cool, but takes a whole lot of time and expertise to set up.Schools need the machines - and you can get a tax break too."
Perhaps there is a greater short-term good and economic incentive to be realized in donating our mothballed PCs to a local school than to re-deploy discarded equipment to solve some GIS problems.
Dimitri Rotow of Manifold (http://www.manifold.net) mentioned that his company has been researching Beowulf clusters and that their team had written algorithms for the Intel Paragon massively parallel supercomputer a few years ago.So when he says, "What one rediscovers is the hassle of working with older machines," one has to listen.His opinion is that there are not many problems in GIS that justify the effort of building a Beowulf supercomputer out of old PCs.But he mentions the one exception: the research environment.Perhaps we should all donate our PCs to ORNL and see what they come up with.
Bill Huber of Quantitative Decisions (http://www.quantdec.com) was also impressed with ORNL's effort, but the skeptic in him said, "do the math..." and his estimate was that it would require about 40 486 processors in a Beowulf cluster to equal the performance of a single modern 1.1 GHz machine.Taking into account RAM, I/O, and disk storage, Huber estimates that you could duplicate the ORNL system by purchasing three modern high-end COTS (Commercial Off The Shelf) PC machines for about US $6000 total.
He then goes on to point out the obvious fact that the ORNL scientists value their time at $0/hour.Yes, I suppose you can do a lot with PhD wizards when you have them for free.Then there is the space and energy issue...just look at the images on the ORNL page and note the floor space requirements, power supply and waste heat issues involved.These are costs only an R & D lab could accept with equanimity.But still Huber isn't throwing out his old PCs in the hopes that they might yet power something interesting.
Albert Janzen cautions that old machines of any value are hard to come by, echoing Rotow's point about "hassles." He also adds that using an unusual processing method for GIS work will come under severe scrutiny by the "old school boys." Maybe that's a significant argument if you have to deal with politics, but it's worthless if you are dealing with open minds.Either the result is correct or it is not; how the problem is solved should be a secondary issue.
Brian Maddox of the United States Geological Survey (http://www.usgs.gov/) was more focused on what GIS work could be done with a Beowulf cluster of surplus machines than on the economics of such an effort.Last year he used 16 machines to test problems using the Clark Urban Growth model (see http://edcwww2.cr.usgs.gov/umap/htmls/pubs.html -- Editor) and also experimented with techniques to reproject and display very large amounts of geospatial data.On the results of last year's work, he is now experimenting with tuning the system to minimize network congestion.
In summary, re-employing your old PCs as a Beowulf supercomputer is very do-able nowadays, and your hard costs in terms of equipment and software are low or zero, but the cost in time involved to learn how to do this and to adapt GIS software to this environment may be significant compared with the short term gains.Longer-term gains have not been proven yet.It's still in the realm of research and Moore's Law continues to make the numbers in the cost-benefit equation like so many moving targets.
Still, wouldn't it be fun to build your own supercomputer?