Flashmob Computing: No Threat To Supercomputing

The "Flashmob 1" gathering of laptop computer users in San Francisco fizzled in its attempt to take a place in the Linpack benchmark of supercomputing, but it scored high in enthusiasm and fun, proving that hundreds of computers can be linked together for some uses.

Jack Dongarra, the University of Tennessee computer scientist who authors the Linpack report asked: "The question one has to ask is: what are you going to do with it?"

In an interview Monday, Dongarra said the informal tying together of hundreds of laptops -- about 700 were interconnected by volunteers at the University of San Francisco over the weekend -- can have a future, for instance, if friends and colleagues want to carry out some advanced calculations.

However, the Flashmob group had aspirations of initially using the installation to perform ambitious calculations normally carried out by massively parallel supercomputers. The "flashmob" idea refers to gatherings that are spontaneously announced over the Web, usually to carry out some offbeat activity or stunt.

Dongarra said the organizers appeared to have set up networking well by using 100 MB Ethernet links and then a higher level network to accommodate the laptops what were lined up in the college's gym. The installation performed 180 billion operations a second -- well below its planned 500 billion operations a second. According to reports from the San Francisco flashmob installation, a few individual laptops failed, causing the installation to slow down.Dongarra indicated that ganging microprocessors together isn't the best vehicle for performing advanced scientific calculations. He noted that supercomputers designed specifically for supercomputing chores -- like advanced Cray Computer machines -- are better suited for traditional supercomputing tasks. The Cray machines and other supercomputers designed specifically for supercomputing are designed from the ground-up to have excellent interconnections between the internal processors and memory -- a feature lacking in multiple microprocessor applications.

Collections of home computers, he added, are unlikely to ever operate efficiently enough for most traditional supercomputing work. The proliferation of many individual processors can lead to software complications, he said.

The most successful ganging of microprocessors to date -- the so-called "Big Mac" installation of 1100 Apple G5 dual processors at Virginia Tech -- ranked high on Dongarra's Linpack report, performing 10.2 trillion operations per second. However, that installation is not well-positioned to carry out some traditional supercomputing operations, Dongarra indicated. The Virginia Tech installation's G5 processors are being upgraded to Xserve G5 servers. Meanwhile, at Virginia Tech, Cal Ribbens, assistant head of computer science, said the smaller more efficient form factor of the new Xserve G5's should help in cooling the installation. He noted that the university has received interest from national scientific laboratories interested in making use of the installation.

The Virginia Tech installation placed third on the most recent version of the Linpack report. In an interview, Ribbens said the computer science staff is working on improving applications for large-scale simulation, system level operations, and fault tolerant protection.