Sun’s R&D spectrum

June 6, 2005—Mention Sun Microsystems to someone, and it’s likely to evoke images of high-powered workstations, dot-com servers, and Java. But a peek inside Sun Microsystems Laboratories reveals a much broader array of emerging technologies and hints at a new Sun rising.

Sun Labs in Menlo Park, Calif., employs some 200 scientists and engineers and spends 0 million to 00 million a year. Its projects include sensors, supercomputers, high-speed networking, optical interconnects, third-generation Web technologies, Java, and more. Its mission: “To solve hard technical problems brought to us by our customers,” says Glenn T. Edens, director of Sun Labs.

Making the switch

For example, Internet switches capable of handling tens to hundreds of terabits of traffic per second today cost millions of dollars and fill entire rooms. But if an ongoing project at Sun Labs is successful, such switches will have dimensions and price tags more like those of PCs. “It’s a high-risk, high-return project. We think it will work, but we don’t actually know,” Edens says.

Ultracheap, high-capacity switches are just one potential application of a technology called proximity communication that Sun announced last fall. Proximity I/O, as it’s also known, can enable processor chips to communicate 60 times faster and with 30 times less energy than is possible using conventional means.

“Proximity allows us to very nicely spread a switch out over a number of chips and have enough bandwidth between them so we can have a distributed switch,” says Robert Drost, a research scientist at Sun Labs. “Proximity” refers to the positioning of two chips extremely close to each other, each with transmitter and receiver circuits. Data is sent across the gap by “capacitive coupling,” which is coupling between charged particles that are at rest. It’s simple in principle, but it’s devilishly difficult to align the chips to tolerances of a few microns.

In proximity I/O, the long communication paths on printed circuit boards with soldered connections and wires are replaced by the tiny, simple interchip gaps. “When processors went from 10 MHz to 3 GHz, they didn’t become 30 times faster, because the bandwidth didn’t increase by 30 times; it increased by two or three times,” Drost says. “[Proximity I/O] will finally realize the potential performance on the chip.”

In addition, he says, very fast interchip communications could reduce the need to have big on-chip caches, freeing up scarce chip real estate for other kinds of processing functions.

Supercomputers

Last July, Sun won a three-year, 0 million contract from the Defense Advanced Research Projects Agency to design a supercomputer with ultrahigh internal bandwidth based on proximity I/O. IBM and Cray each won awards for designs based on different principles.

Drost says the supercomputer will be “massively parallel,” with hundreds of thousands of threads executing in parallel. It will excel at problems that require a lot of interprocessor communications, such as database searches, scientific simulations, and signal processing. If Sun wins approval to build a working machine in the next phase, one or more prototypes could be installed by 2009, Drost says. Those machines would run at sustained speeds of 1 quadrillion floating-point operations per second (PFLOPS), about 15 times faster than the fastest supercomputer today, and be scalable to 4PFLOPS.

Sun Labs is working on computers at the other end of the spectrum as well, and it claims to have developed the world’s smallest secure Web server. Code-named Sizzle, the server is the size of a quarter and is intended to go inside home appliances, personal medical devices, sensors, and the like. It’s a battery-powered, wireless device with an 8-bit processor, 128 KB of flash memory, and 4 KB of RAM.

Others have built tiny Web servers, but what distinguishes Sizzle is its use of elliptic-curve cryptography (ECC), which is more efficient than RSA cryptography and hence more suitable for compute-challenged processors.

Users of the industry-standard RSA have moved to 1,024-bit encryption keys and will eventually have to move to 2,048 bits to ensure that the keys aren’t broken. Every doubling of key length requires an increase of computer power by a factor of eight.

But ECC at comparable strengths is 10 times faster than 1,024-bit RSA keys and 38 times faster than 2,048-bit RSA keys, says Vipul Gupta, a senior engineer at Sun Labs. “The next generation of Internet devices, such as sensors, is expected to be even less capable than phones, and these devices just don’t have the horsepower for RSA,” he says.

Gupta has worked with the Internet Engineering Task Force to get ECC integrated into the Secure Sockets Layer encryption protocol, just as RSA has been integrated with it. Now, he says, developers can write software that interoperates with Sizzle as easily as with any other secure server. Applications include battlefield sensors, personal medical devices, and radio frequency identification tags for confidential situations.

Gupta says ECC may find applications in large servers as well. A big e-commerce company such as Amazon.com could get by with a quarter to half as many servers if it used ECC rather than RSA, he says.

Security isn’t the only thing Sun Labs is trying to get to work on tiny computers. Its Project Epsilon is based on the belief that real applications of sensor networks are scarce because sensors communicate unreliably and are hard to configure and program.

“We are working on how to program these things,” Edens says. “And we’re working on radio protocols, because IP was developed with no thought to power savings, and we are working on how to upgrade software if you have 10,000 of these.”

The answer to some of these questions is Java, Sun says. Java will bring interoperability and ease of code migration, says Randy Smith, a principal investigator at Sun. But it’s not easy, he acknowledges. The key is getting Java to run on bare metal—no operating system. “It’s a shoehorning thing,” Smith says.