April 6, 2004

FlashMob I proves points but misses goal

Author: Chris Preimesberger

SAN FRANCISCO -- We came, we saw, we didn't quite conquer. But we made a few points and had fun doing it.

FlashMob I took place on a typically foggy April Saturday at the University of San Francisco's Koret Center gym. Exactly 700 computers of various shapes and sizes -- donated by individuals and several organizations -- were gathered in the auditorium and interconnected via a homegrown local area network to in effect try to become one of the world's 500 most powerful supercomputers. Surprisingly, it wasn't the huge tangle of wires many people thought it would be; all the networking hardware was kept well under control by all the event "hub captains."

Orderly, well-secured event

FlashMob I was an orderly, well-secured event that proved that: a) such a large internetworking of computers could indeed be organized in only five weeks, b) so many computers could be connected and made to work as one in a short six-hour period of time; and c) 700 computers could be delivered, set up, booted up on the network, and returned to their owners intact with few security problems in one afternoon.

USF professors Peter Pacheco (left) and David Wolber are interviewed by a local television station during FlashMob I.

The original goal was to put together enough computers to create one of the world's most powerful supercomputers. That didn't happen; the administrators from the USF computer science department found that they needed more than one afternoon to approach that lofty goal.

The goal had been to connect enough machines to produce about 430 gigaflops of throughput as measured by the Linpack Benchmark. Since computer No. 500 in last year's Top 500 Supercomputers listing rated 403 gigaflops, organizers of FlashMob I figured it would take that and a bit more to inch into this year's listing. As it turned out, 669 of the 700 computers actually worked and were linked. One large bank of 256 was connected to produce a high of 187 gigaflops of throughput for a few minutes, and 150 gigaflops steadily for about 40 minutes. A node went down about three-quarters of the way through the computation, however, ending the best benchmark attempt.

Our NewsForge ThinkPad, running a Knoppix derivative called Morphix from a USF-developed networking CD (as did all the computers used that day), was computer No. 669. We had to register, tag the computer, obtain two credential cards (one for the owner, one for the computer -- that one had a photo of yours truly on it), get escorted to and from the event floor, and leave the computer for six hours while the benchmark was attempted. Picking up the computer at 6 p.m. was also a marvel of security; no fewer than six people checked us as we reclaimed our hardware. No equipment was "misplaced."

The main problem in not reaching the goal, said event volunteer, IBM programmer, and hub captain Damien Eversmann, was that there was a "flaky network card in one of the nodes. It took a long time to figure out which one it was. So we were only able to get one section (256 machines) up and running. If we had had more time, we could have identified the problem machine and replaced it."

Ran out of time

As it was, USF promised to have all the machines ready for pickup at 6 p.m., so the event had to end at that time. Organizers tried several times to reach the benchmark, each attempt taking about an hour to get the machines booted and running on the network.

Network monitors make an attempt to run the Linpack Benchmark during FlashMob I.

"We had good performance (187 gigaflops) on only about one-third of computers we had running," said David Wolber, chair of the USF computer science department. "If we had had most or all of the machines running, we could have reached our goal. We just ran out of time; it took quite a while to get all the machines booted and rebooted in order to synchronize."

But Wolber and fellow professors Peter Pacheco and Chris Brooks were all happy with the event anyhow.

"Just to think we could pull this off, with only five weeks' preparation, is amazing," Brooks said. "We also attained some good public awareness about power computing. It was a lot of work -- we wrote all the networking software right here in the department -- but it was worth it."

The idea for FlashMob I grew out of Professor Pat Miller's "do-it-yourself" computer-building class at USF. "We just thought, hey, why not take this on a big scale and make a big event out of it?" Brooks said.

Wolber said that preparation will start immediately for FlashMob II next spring. "Of course we're going to do it again," he said. "We learned a lot this time. We're already looking forward to next year."

Click Here!