CNET también está disponible en español.

Ir a español

Don't show this again

Tech Industry

Sun's x86 strategist steers straight ahead

Sun's John Fowler navigates a tricky transition as he guns for growth in x86 servers this year.

John Fowler holds half of Sun Microsystems' server future in his hands.

As executive vice president of the Network Systems Group, one of two server groups at the company, he leads Sun's belated but now vital push to sell servers using x86 processors. Only a small fraction of Sun's revenue comes from x86 servers, but the overall market has been growing consistently for years--and Sun craves revenue growth.

Sun's mainstay business, servers using the company's own UltraSparc processors, has been hit hard by IBM's charge into the market and the arrival of Linux. Now x86 servers are key to Sun's attempt to outflank its rivals. Sun has made some gains, ranking sixth in the x86 server market. The goal is to be No. 4 by the end of this year.

Sun's x86 push began in 2002 with undistinguished Intel-based machines running the Linux operating system. It picked up some steam in 2004 when Sun released its first servers based on Advanced Micro Devices' Opteron. But the systems were designed outside of Sun and lacked certain features, such as redundant power supplies, that were demanded by businesses. In October, Sun began selling its own Opteron servers designs, code named Galaxy and designed by Sun co-founder Andy Bechtolsheim.

I'm not a basher. I may talk like a wild man, but inside the team we take a different approach.

Fowler, 45, is a somewhat unlikely pick for a hardware executive. His background is in software, including two years as . When Sun's software chief, Jonathan Schwartz, was elevated to president in 2004, he named Fowler to the x86 server post.

Fowler discussed a range of subjects with CNET, including his belief that Opteron will lead Xeon for years.

Q: How is 2006 going to be different than last year for Sun?
Fowler: I won't be explaining AMD anymore. We spent a lot of energy in 2005 explaining what it was about. In the enterprise part of the customer base, they've gotten that message. The second thing I won't be doing is answering the question, "You guys are in the x86 business?" Or, "Are you serious?" Customers have gotten their hands around the idea that this is something we're really doing.

But only a few months ago (Sun President) Jonathan Schwartz and others were saying customers still didn't know about your x86 servers.
Fowler: The Galaxy launch was huge. We measure penetration into customer accounts and awareness factors. We're getting there. In SMB (small and medium businesses), in particular, we're not well known. That's part of a longer-term thing to work on. But in core Fortune 200 we're known.

We'll have bunch of new products in 2006. The first Galaxy products were the beginning of the family. There is a set of common architectural elements you'll see carry over into more rack-mount products and blades.

Your designs use higher speed-grade Opteron processors (models that run faster but consume 120 watts compared to 95 watts for standard Opterons). Do you expect to stay one speed-grade ahead of the pack?
Fowler: Absolutely. AMD will continue to produce some kind of higher speed-grade. We didn't do it just to get one speed-grade ahead in one generation. In some applications, adding speed-grades makes a lot of difference. If you have a very expensive software license, having an additional speed-grade makes you that much more effective in software.

On classic Web services workload, Niagara will spank any Opteron.

Are you going to launch the eight-processor system and the blade system separately?
Fowler: We're still trying to figure out what are the best ways to launch the products.

High-end x86 servers are rare. Do you have to educate customers that eight-socket big-iron servers are worth buying?
Fowler: Absolutely. People have to go a life cycle in purchasing. Initially, it'll be by people in technical computing and people who are classic early adopters building Web services infrastructure as opposed to people doing (enterprise resource planning) and (decision support systems). People aren't going to get up in the morning and say "I'm going to run my Oracle database on an eight-socket system."

You say you monitor your customers. What fraction are using Galaxy?
Fowler: I don't have a handle on that. The fraction is really small compared to total number of enterprise customers, which is in the thousands.

What's the Solaris/Linux mix, and what's the goal for 2006?
Fowler: The attach rate of Solaris is very difficult to get to the bottom of. Last year, we moved to preinstalled Solaris 10 on all my machines. Every one that goes out the door, as long as it's not diskless, includes Solaris 10. What happens after that is hard to figure out. Customers can install their own operating system.

But we just introduced the Solaris 10 update 1. It includes a registration automatic-update facility. We're going to start getting statistically relevant data on how many people are running Solaris and what updates they're dong.

In the past, we quoted 20 to 25 percent in terms of Solaris usage. Now Solaris usage is way up.

And do you have a goal for 2006?
Fowler: No, we don't have specific goals for Solaris' percentage. We're looking at the overall picture of unit volume. If we sold a million servers and the percentage of Solaris isn't very high, it would still be good for Sun.

But not better for profit margins?
Fowler: There are a number of reasons Solaris is important to us. We can bundle service with hardware. We wrote it, so we can fix it. Recently we had a situation with Niagara (Sun's new UltraSparc T1 processor) where a customer noticed something odd. In less than 24 hours, we found and fixed the issue. You can't do that if you didn't write the operating system.

With Solaris running on eight-socket x86 servers, are you competing against your company's Sparc product lines?
Fowler: The product lines overlap today. The (UltraSparc-based) V120, V240 and V440 potentially overlap with the (Opteron-based) V40z. The decision we made quite some time ago was that we were going to go and make a server product line that would solve whatever problems people had.

If you look at the newer UltraSparc stuff like IV+ and Niagara, the workloads they're best at are different from what Opteron is good at. On classic Web services workload, Niagara will spank any Opteron. On general-purpose workload, Opteron is faster than Niagara. The sales force figures out what is the best way to solve problem then uses whatever technology works. Having Sparc in the back end a V20z and V40z in the front end is a common engagement now. People are using Niagara with Galaxy.

Who are you up against in customer accounts?

Fowler: The companies you see over and over are the big three: Dell, HP and IBM. In financial services, a selection of customers strategically chose to go with Opteron and they're no longer purchasing Intel. In those accounts, it's invariably us and HP. In more general x86, it's constantly the big three. A year ago, I wasn't necessarily even getting invited to compete.

There are customers who buy Opteron only?
Fowler: There are companies that have chosen to make future purchases be Opteron only for the entire infrastructure of x86, and they're quite large. It's not that surprising. Years ago, they decided to be Intel only. The way it works is: you can buy something else, but you have to go through an exception process in purchasing.

Do you think the changes out of Intel will be significant this year and next year? They have "performance per watt" emblazoned on their marketing materials these days.
Fowler: They'll use the advertising side to promulgate a message. On the technology side, they're going to make an incremental improvement, but the challenge is the front-side bus architecture (the data pathway that links the processor to the computer's memory system and other components) is something that doesn't change. You can tweak the front-side bus speed and do other things, but what we've discovered: Niagara has on-board memory controllers. If you can tackle memory latency, almost everything runs faster. Memory latency is not going to change (with Intel). I don't see the competitive landscape changing a lot between Intel and AMD.

AMD's transition to new 65-nanometer manufacturing technology is slower than Intel's, which has price consequences. Are you concerned?
Fowler: Intel has the world's best manufacturing. At the end of the day, the architecture and capability of the technology is as important or more important than the nanometers of the process you can put it in.

One advantage AMD has with the onboard memory controllers is that they can get away with smaller (memory) cache sizes. They are achieving their great performance with 1MB caches. One of the side effects is they don't actually need the same number of chips-per-wafer yield to get to the same price points because they're not trying to put 8MB of cache on a chip. Intel needs 65 nanometer parts because they have to put these put on these huge caches.

Sun's core customers are big companies. To compete against HP's x86 strength, you have to go after smaller companies.
Fowler: We have a pretty good channel structure and partner structure to reach the companies with 1,000 employees and up, or 500 and up outside the United States. As we continue to grow we will extend the reach.

We have healthy respect for our competitors. I'm not a basher. I may talk like a wild man, but inside the team we take a different approach. For example, in our product strategy, we built products around not just market analysis and size, but mapped it against emerging technologies and Sun technologies like Solaris. Our engineering strategy is aligned with areas we think are going to grow. A traditional IDC analysis shows the eight-socket x86 market is not very big. When we looked at technology trends and Solaris and virtualization, we concluded that it's going to be huge.

You like AMD chips and say they're good for general-purpose computing. But your counterpart at IBM, Susan Whitney, sees things differently, saying Opteron is only good for the narrower segment of high-performance technical computing. Why?
Fowler: If you don't have an Opteron line, that's an important statement. If you look for example at SQL Server benchmarks (or) directory services on Opteron, the commercial benchmarks are compelling, and in the commercial space people care endlessly about performance per watt. I see a lot if interest in enterprise and outside high-performance computing. I think it's my word against hers.