In high-end Unix circles, Non-Uniform Memory Access (NUMA) architecture is being touted as the design of the future, the entry point to high performance for thousands of Unix users and the edge that Unix suppliers need to shut off inroads that Windows NT might make into their market. This is high promise for an unproven technology, but NUMA may just be all of those things.
"NUMA is going to be extremely important for high-end systems," says Brian Richardson, program director of Meta Group in Stamford, CT. "It represents a hybrid architecture with the best of symmetric multiprocessing [SMP] and massively parallel processing [MPP] architectures."
Adds Scott MacGregor, senior vice president of products for SCO in Santa Cruz, CA, "It's cheaper than SMP but delivers the same performance, versus a clustered approach in which you give up a lot of performance."
NUMA achieves its high performance level by having all processors share a consistent global memory but allowing the memory access time to vary (to be nonuniform), based on the memory location. Access to close memory--that is packaged on the same board as the processor, such as registers, cache and local memory--is much faster than distant memory--that is packaged on other processor boards. In traditional SMP designs, memory access is uniformly governed by the speed of distant access.
MPP designs, on the other hand, move data simultaneously along parallel pathways, providing lower latency and less bus hogging. However, MPP requires massive reprogramming of applications to take advantage of the design--an expensive and often impractical trade-off. Since NUMA memory appears as conventional SMP shared memory to an application programmer, NUMA systems can run existing SMP applications without modification. In this way, NUMA bridges the gap between SMP and MPP. In the judgment of some analysts, it will become the dominant design technique for the coming generation of SMP systems.
An Edge for Unix
NUMA design also can take advantage of Intel's Standard High Volume (SHV) "quad" servers, which are boards consisting of four 200MHz Pentium Pro processors with 512K of cache per processor, up to 4GB of memory and dual Protocol Control Information (PCI) I/O channels. Using these Intel boards, NUMA architecture is expected to reach scalability up to hundreds of processors, giving Unix-based systems the ability to achieve performance levels previously unattainable using commodity hardware. That's expected to bring high-performance servers into a price range affordable for many more users.
Unlike high-end Unix implementations, Windows NT has not been designed to take advantage of NUMA and won't be until after year 2000, according to Richardson. "Unix will be more appropriate for high-end systems that require SMP scalability significantly beyond four-processor systems," he says. "Unix will use the SHV building blocks to construct more-scalable systems. This will give Unix a more established position as a high-end operating system platform versus NT. NT will come to dominate the one-to-four-processor space."
Adopting this technology is no simple matter, according to MacGregor of SCO. "It's a fair amount of work to take advantage of NUMA," he says. "You have to have an operating system that's sufficiently modular to be able to parcel off the different parts of applications across processors and understand the hardware architecture. Microsoft has a portability layer on Windows NT, but it doesn't work particularly well on NUMA machines."
Microsoft could reprogram NT to take advantage of NUMA now, but it won't because of the company's high-volume focus, says Bob Robinson, product marketing manager for Sequent Computer Systems in Beaverton, OR. "From their perspective, the number one requirement is to sell lots of licenses. There are millions of PCs, and hundreds of thousands of departmental servers. At the enterprise level, there is a magnitude less."
The high performance levels provided by NUMA will be important immediately, both for large database access at the enterprise level and for higher-end departmental application servers. Vendors also will be marketing their NUMA systems to those needing World Wide Web servers capable of withstanding the onslaught of millions of browsers demanding access. "We're talking to customers who want 15 to 16 terabytes [of data storage] in 1997, and it's just going to grow from there," says Robinson. "You have to have a very high-performance system to handle that."
Companies developing NUMA machines include Sequent, Data General, Pyramid Technology and Cray (now part of Silicon Graphics). Data General plans to market its NUMA systems to the mainstream business corporate market. "We think it will replace many RISC and mainframe products," says Phil Gerskovich, vice president of DG's NUMALiine business unit. "You will get much more performance at lower cost." Another advantage of NUMA, Gerskovich says, is that users can buy a NUMA-enabled four-processor node and expand its capacity later by adding processor boards. "We think people at the departmental level will buy those even if they don't have the need for the large systems initially."
Richardson of Meta Group expects a NUMA machine from Sequent in late 1996 and from Data General in early 1997. The vendors are differentiating their products in part based on the transport layer. Sequent is employing a chip from Vitesse Semiconductor of Camarillo, CA, for the physical transport and has developed a custom application-specific integrated circuit (ASIC) to handle the transport protocol. Robinson says Sequent's NUMA-Q system architecture will enable its first machine to scale up to 252 processors.
Data General plans a 32-processor machine that employs a transport solution from Oslo, Norway-based Dolphin Interconnect Solution A/S. Dolphin's implementation of the IEEE Scalable Coherent Interface (SCI) effectively bridges traditional bus architectures, such as PCI, with a high-speed, point-to-point interconnect protocol.
SCO's UnixWare is being rewritten for NUMA, but the first NUMA machines are expected to run vendor-specific Unix versions, namely Sequent's Dynix and Data General's DG-UX. Data General and SCO have agreed to jointly develop the NUMA version of UnixWare, and DG's machines will also run UnixWare.
"A large number of vendors will be selling NUMA architecture systems, but only a relatively small handful of vendors will have the expertise to develop them," says Richardson. "I expect to see a lot of OEM deals with cross-licensing." In a memo on the subject, Meta Group predicts that dominant systems vendors Digital, HP, IBM and Sun, who have large investments in RISC processing, will be among the last to field true NUMA systems, relying instead on MPP and conventional clusters in the meantime.
According to statistics from the United States Department of Labor, most of the growth in the U.S. economy these days is coming from small businesses. For many open systems professionals, this means that the route to growth and success may not only be up a corporate ladder but out on one's own. The emergence of one new Silicon Valley startup company reveals much about the bold move from corporate to entrepreneurial life.
While Microsoft labored over and launched the expensive, gaudy Windows 95 campaign, Kim Polese was busy at the grassroots level selling the concept of Java to software developers, the press and Wall Street for her employer, Sun Microsystems. Java has since become the hottest programming language for the Net. Sun hopes that JavaOS and the HotJava browser and class libraries, still in development, will produce the next-generation desktop, usurping the power of Windows.
Today, Kim Polese is president and CEO of Marimba, Inc., based in Palo Alto, CA. This fledgling application tools company is building on what she sees as a "clean slate" platform for the Internet: Java. The company's first products are due to ship this fall.
Polese spent seven years at Sun. During her last three years there, she was the product manager who moved Java from being a narrowly conceived language for specialized computing devices to a red-hot vendor-neutral platform. She recalls putting the words open systems on a white board during a discussion of the future of development; that was a major part of the shared vision that guided the project. In 1993, Polese and what was to become the Java team persuaded Sun to broaden the focus of James Gosling's project code-named Oak by writing the business plan for Java. This included the name Java itself, the positioning and plans for moving it into the marketplace.
Toward the end of 1995, Polese convened three other members of the original Java development team at Sun: Arthur van Hoff, Jonathan Payne and Sami Shaio. This foursome agreed to found what they ultimately called Marimba, to provide next-generation technologies for deploying network-aware Java applications. The group saw opportunities to build on top of Java's success--what was missing were the vendor-neutral tools and applications.
Polese left Sun because she had accomplished what she'd set out to do: "to get the technology out to the market and make it a standard," she says. By September 1995 she and the others were eager to press on and implement their vision.
Polese, van Hoff and Shaio resigned from Sun by telling Scott McNealy, Sun chairman, personally of their decision. (Payne had left Sun earlier.) "We quit on a Friday and started the company on Saturday," Polese recalls. "Scott was surprised, but he recovered quickly." Van Hoff and Polese insisted to McNealy that they wanted to maintain a good business relationship with Sun and solidify Java as a standard. SunSoft is at work on Java tools, and the JavaSoft subsidiary is set to deliver the operating system for new Internet devices. Marimba intends to complement these efforts in the marketplace.
Some Sun insiders seem to worry that Marimba is perceived as having taken the Java brain trust, leaving JavaSoft with the suits. Regardless of how true this may be, Sun must deal with the independent software vendor (ISV) community Java has created--a model with which many platform vendors have had a difficult time in the past. Treating ISVs well builds strength. Shunning them as competitors or trying to outdo them only dilutes a platform's power. Polese believes that McNealy understands the value of a strong ISV community for Java. One indication of this may be that JavaSoft will publish APIs for developers soon, based partly on work done by ISV partners Macromedia and Adobe.
Marimba describes its mission as being to "dramatically improve the experience of both end users and developers on the Web." There are a lot of challenges to overcome in terms of usability of the Web and its content. Marimba will address them, though the team is not ready to be more specific about what shape its products will take. Observers assume that at least one will be object class libraries.
The partners currently are working out financial backing, product positioning and other basic questions. Funding, and how to accept it without giving away power, is one of the biggest issues facing Marimba now. Venture capitalists are all but driving around Silicon Valley with truckloads of money, looking for Internet-related businesses to invest in. Marimba is in the right place at the right time. "We are beating venture capitalists off with a stick," Polese says.
She insists that, in auditioning venture capitalists, money is far down her list; strategic help on building the company and forming partnerships is more important. None of the Marimba team has built a company from the ground up, so they are looking for a venture firm that specializes in strategic planning and recruiting. Growth in Marimba will start with key hires such as vice presidents of engineering and sales and a CFO.
Marc Canter, founder of Macromedia, is a similarly driven entrepreneur and friend of Polese. He feels that he gave too much power to the backers of his own venture and recommends that Marimba get at least three venture capitalists who will work with the founders, rather than being a "Roman Senate conspiring behind your back." In addition to product positioning and venture backing, Marimba is considering alliances and partnerships with a variety of hardware and software companies.
If the founders don't yield to what will be a financial temptation to sell out to a single Internet platform vendor such as Netscape, Sun or Microsoft, they may become a model for software companies in the age of the object class library. They have no hardware or operating system to sell, only implementations of platform independence at a time when the marketplace and Wall Street are finally ready to absorb the concept.
Polese laughs when asked how life in a large company differs from her life now. "The last few years at Sun felt like a startup," she says. "The Java team was not part of the Sparc operation. We worked startup hours, and I had total responsibility for about 10 different jobs." What she is doing now "feels totally natural and in fact less stressful than Sun, because I'm more in control. I've always known I was meant to be an entrepreneur. It was only a matter of time. It had to be the right people, right product, right time in the market. The most important thing to me in this little company is to have a big adventure. Even if it all goes crazy and Java doesn't turn out the way we all think it will, if I learn something and have fun it will be worthwhile."
Of course, the industry, and those who watch and invest in it, are expecting much more from Marimba.