FlashMob Supercomputing Event at University of San Francisco
What has three full-sized basketball courts, six regulation hoops, and 180 gigaflops?
The University of San Francisco’s Koret Gym met that description during
this past Saturday’s FlashMob I supercomputing demonstration. Sneaker-clad
onlookers clutched their gym bags and gawked in amazement as volunteers loaned
their 700 personal computers to be linked together for the day, transforming
the gym for about 6 hours into a world-class, albeit temporary supercomputer.
FlashMob I demonstrated, in both technical and logistical terms, the potential
of ephemeral computer clusters for supercomputing. Of the 700 volunteered machines,
669 were eventually hooked up to the network, to be run in differing configurations.
Benchmarks were run throughout the gym’s short career as a supercomputer
center, using LINPACK, a familiar benchmarking software in supercomputing circles.
While the temporary supercomputer would not be added to the "Top 500"
list, measurements indicated that it was within range of some of the top supercomputers
in the world.
But more to the point, the event’s organizers say, FlashMob I shows the
validity of the notion of a temporary supercomputer. Depending upon need, such
cooperative events could be scaled and targeted to address particular problems.
"The purpose of this new model for computing is to match up people with
science, so that they can work on the common problems they are interested in,"
says Greg Benson, assistant professor on the Computer Science faculty at USF.
Benson, who first linked the term "FlashMob" with computing, noted,
"We had an incredible mix of people today: high school students, USF students,
and top scientists in supercomputing…" He remarked about the difference
between the exclusivity of most supercomputing centers and the inclusiveness
of a FlashMob-style supercomputer: "The current model for supercomputing
is that if you as an ordinary citizen wanted to run something on a supercomputer,
it would be very difficult for you to do that… The whole point here is,
if you get enough people who have a common goal, or a common concern, to get
together and work on the problem that is interesting them, they will have a
lot of choice."
"It’s part social engineering, part computer science," says
Pat Miller, a computer scientist and researcher at Lawrence Livermore National
Laboratory and a lecturer at USF, comparing the project to a barn raising. "Collectively
we can do bigger things." Miller was pleased with the benchmarks: "We
got a really world-class number today; 180 gigaflops [180 billion floating point
operations per second; with 256 computers connected, even though one node failed
at 75 percent completion]." Pointing out that large, high-performance computational
clusters may typically cost half a million dollars, he says that temporary clusters
could make it far less costly to approach supercomputing. Unlike permanent clusters,
FlashMob d'es not involve purchasing and maintaining costly equipment. "Here,
the idea is when you need it, you just create it, simply with the things you
have onhand."
FlashMob organizers consider their best number of the day to be 77 Gflops,
with 150 of the computers connected and the benchmarking running successfully
to completion. The participating computers of FlashMob I were connected by software
designed by students and faculty at USF. The open source software is available
for similar computing projects. Besides the volunteers’ computers, logistical
support, facilities, cabling, and other support and expertise needed for the
project came from USF, and industry sponsors Foundry Networks, Myricom, and
HP provided additional financial and technical support. www.flashmobcomputing.org
[Editor’s Note: Greg Benson will offer his reflections
on FlashMob I in Thursday’s IT Trends e-newsletter. To sign up for the
newsletter, click here]