May 16, 2002
Industry adopts UW researchers’ innovation that promises huge boost in speed and efficiency for high-traffic computer chips
The speed and efficiency of computer network and database servers could increase as much as 400 percent because of an idea developed by two University of Washington computer scientists that is reaching mainstream computing.
The technique, known as simultaneous multithreading, was first proposed in the mid-1990s by Susan Eggers and Hank Levy, UW professors in the Department of Computer Science & Engineering. Compaq Computers was the first to announce a processor using that idea on its Alpha computer, but that line was discontinued before Compaq’s merger with Hewlett-Packard. However, Sun Microsystems announced in December that it would produce microprocessors using simultaneous multithreading and Intel Corp. made a similar announcement shortly afterward. Other companies appear likely to follow suit.
The new concept will have a major impact on high-end computer users, such as Internet service providers and businesses with server-administered databases and networks, Eggers and Levy predict.
“At computer conferences, presenters talk about new ideas that will boost speed by 15 or 20 percent, and people get really excited,” Eggers said. “We’re talking about an increase of 300 to 400 percent.”
Ideas that fuel momentous forward leaps in computing come along about once every 10 years, Levy added. “This, we think, is one of those ideas.”
Burton Smith, chief scientist at world supercomputing leader Cray Inc. and a pioneer of multithreading technology, agreed, adding that he expects simultaneous multithreading, or SMT, to become an industry standard.
“SMT fits multithreading into the architecture that microprocessor manufacturers are using in a relatively painless way,” said Smith, who was the first person to use multithreading in the central processing unit of a computer. “It provides them with a much better use of the hardware and much better performance as well, with minimal impacts on the way they do business. It’s quite a nice thing.”
The concept of multithreading was first proposed in the 1960s and is used in supercomputers. Eggers and Levy’s technique shows how the idea can be applied to mass-market computing. The key to SMT involves making better use of a chip’s computing power.
Modern computers are capable of doing a lot of work simultaneously. For example, some processors can perform 10 arithmetic operations (like adding or subtracting) every clock tick, which is every billionth of a second on a one-gigahertz processor. In practice, however, they do only two or three operations per tick, wasting the potential for much higher performance. That’s because modern computers work on only one program, or thread, at a time, and there’s a limit to the number of arithmetic operations a single program can present to the computer in one clock tick.
Imagine a chef who is cooking dinner for four people, standing in front of a four-burner cooktop. If the chef prepares one meal at a time using one frypan on one burner, he’s wasting three-fourths of his potential. On the other hand, if he places one frypan on each of the four burners and prepares the meals all at once, he’ll be finished in about one-quarter the time. SMT is like the efficient chef — it takes instructions from many programs and loads them all into the computer at once, using the full arithmetic potential of the computer to process many programs in parallel, like the chef who’s sauteing in all the frypans simultaneously. This is particularly useful for Web servers, which are trying to cook thousands or millions of “meals” at the same time.
The trick lies in keeping close track of each thread. To do that, some additional hardware is required so that the state of each thread can be stored and updated.
That extra hardware requirement highlights the advantages of SMT, according to Eggers. The additional circuitry for a chip that can run four threads at once is estimated at about 10 percent. But the boost in performance far outstrips the hardware need because when one thread is stalled, waiting for data, the other threads keep running. That’s what makes SMT so advantageous for database and Web servers, which generally process user requests as separate threads. According to Eggers and Levy’s research, an eight-thread SMT chip can run database software three times faster than a normal chip. For Web-server software, performance jumps fourfold. And that’s with unmodified software, Eggers said. Fine-tuning programs with SMT specifically in mind could make things even faster.
“It gives you a huge bang for your buck,” she said.
Ultimately, Cray’s Smith predicts, multithreading of one sort or another will be needed for all types of computing, from personal computing to the very highest performance computing we do.
“The circuitry is getting so much faster and transistors are getting less expensive,” he said. “To take advantage of the opportunities that presents, multithreading is the only way to go in the long run.”
In that sense, SMT is a start.
“It’s a beginning for the mainstream of computing in exploring a whole new kind of architecture,” he said. “It will become standard and there will be further evolution of these ideas. It opens all sorts of possibilities.”
###
For more information, contact Levy at (206) 543-9204 or levy@cs.washington.edu, or Eggers at (206) 543-2118 or eggers@cs.washington.edu. Additional information about simultaneous multithreading, including links to academic papers that proposed and developed the concept, can be found on the Web at www.cs.washington.edu/research/smt/.
