The NCSA Symera Distributed System, being developed by a team led by Briand Sanderson and Rick Vestal aims to not only soak up idle computing cycles but also, because it is geared toward parallel programs, to pack the processing punch of multiple-processor supercomputers. Symera won't compete with the high-powered systems in speed, but it is preparing to go head-to-head in user satisfaction and data output.

Symera won't just be for down time," says Vestal. "We think of a company or institution buying 25 workstations, placing them in a room, and using them just for computing because that will be cheaper than buying a supercomputer."

One reason the creators of this upstart are confident it will compete with supercomputers is that Symera is based on Microsoft's DCOM technology and, hence, will work optimally on the increasingly popular Windows-NT platform. DCOM, which is better known for its implementation as ActiveX, is a spec sheet of instructions on how self-contained pieces of a program, called objects, talk to each other. The communication protocols are similar to the cables connecting speakers and tuner in a CD player in that they link together the pieces so that they may function as a unit.

Some object-oriented programming languages have clunky communications protocols. For instance, to use UNIX sockets (network communication paths) written in C in an NT environment, programmers have to write programs that pack and unpack byte streams of data as they leave and then arrive at machines in a cluster. ActiveX is simpler. It creates objects on other machines so that the programmer needs know nothing about how the network operates between the machines. "It's a piece of cake to learn and use," says Vestal, who didn't have to write a single line of communications code for Symera, which inherited ActiveX's model. Symera also inherited ActiveX's strong security features. Thus, while other clustering programs are struggling to keep out hackers, Symera is already bolted shut. Says Vestal, "Joe Schmo can't break into our system and run an object that will crash the hard drive."

Symera also emphasizes parallel, multithreaded applications. Multithreading decomposes an application into sections that can be run simultaneously, either on the same processor or among several processors, which is akin to traditional parallel processing. Symera parcels these chunks of program, or drones, among several machines, taking advantage of efficiencies of scale. Theoretically a Symera job could run on thousands of processors (assuming the application exhibited a similar degree of parallelism). However, communications bottlenecks are likely to keep actual processor counts much lower.

Many features in Symera are undeniably similar to those in the University of Wisconsin's Condor system, which will also be implemented at NCSA. The Symera team is coordinating its NT efforts with the Condor team so that the two systems will be interoperable. Still, Sanderson credits the inspiration for Symera's parallel computing architecture to parallel virtual machines (PVM). These collections of libraries enable UNIX workstations to talk to one another as they coordinate the intricate steps in a parallel application. PVM systems still require researchers to manually transmit each portion of their application to whichever machine in a cluster it will be running on. That necessitates obtaining access to each machine and establishing communications protocols. Symera eliminates such manual intermediaries.

Further streamlining communication in Symera are applications templates called wizards--handy shortcuts that spit out code--that establish the communications protocols. Researchers merely fill in a few functions and allow the system's drones to do the rest. The wizard written by the Symera team currently works only for applications written in C++. Eventually, however, other wizards may conjure code for all the major programming languagesÑVisual Basic, Java, C, and Fortran. Researchers would not even need to fill in functions if Symera were strictly for running batch jobs, but it is not.

As Sanderson proudly points out, "Symera lets you steer and interact with the computing resources." Jobs don't enter a computing black box that spits out final computations. Symera will soon have programs for tracking jobs so that researchers can stop or change them as they would if the jobs were running on their desktop machines. Says Sanderson: "If you see something in your simulation you think is interesting, you stop and examine it, or you change one of the parameters to see what affect it has. That's interactive computing, what we call the Immediate Mode. Lots of clustering systems promise high throughput (computing lots of data) but interactivity is why Symera will compete with supercomputers."

The Windows-NT beta version of Symera is expected in July. (Later versions will support UNIX.) Any researchers willing to test the system in exchange for free NT cycles should contact Briand Sanderson at briand@ncsa.uiuc.edu.