Contrarian Minds: David Yen

What comes after throughput computing?

By Al Riske

01.Jul.04--David Yen doesn't consider himself a contrarian. It's not respectful. And yet the head of Sun's Scalable Systems group has been bucking some widely accepted industry trends, and doing so quite aggressively.

While he recognizes that current processor technology is "good enough" for many tasks -- and in some cases is already overkill -- Yen believes emerging challenges will require much different solutions.

So instead of designing ever-more complex microprocessors and chasing after higher and higher clock speeds, Yen's team has focused on just the opposite: simpler designs that are, in some respects, actually slower. But here's the twist: They get more work done.

It's hard to imagine a more contrarian approach, but it works because it addresses the realities of network computing today: It's not about doing one thing really fast; it's about doing a lot of things at once.

The multicore, multithreaded design of Sun's new microprocessors -- introduced just a year after we announced the overarching throughput computing concept -- makes them extremely good at juggling computing tasks.

Well, to be accurate, it's not just the design of the processors.

"To realize this whole vision, we believe it takes a number of essential components," Yen says. "You have to have the right processor design, the right memory bandwidth to feed the hungry processors, and the right commanding software, the Solaris Operating System, to efficiently handle all the thread scheduling and resource allocation."

All of which gives Sun a distinct advantage in the new paradigm of throughput computing and its logical extension, throughput networking.

Look at how much e-mail we send and receive each day or the hundreds of millions of queries search engines handle around the clock.

Brute force isn't the best answer, Yen says. You have to address the unique nature of network computing workloads.

A network-facing processor has two jobs to do. First, it has to manage the flow of data on the network. Second, it has to process the data. So there's a lot of ping-ponging between tasks.

"In the old days, this could be viewed as flexibility -- it's a general purpose processor and I can ask it to do anything I want," Yen says. "It gives me flexibility but is not necessarily the most efficient way to do it."

One of the problems is that, under "packet pressure" from all the IP devices and sub-IP devices on the Net today, an inundated processor will start to drop packets. Dropped packets have to be resent, which creates even more congestion on a network.

"We are already aggressively pursuing this multicore/multithread thing, so now we can afford to specialize some of the cores to more efficiently execute certain operations without continually relying on software," Yen says. "Anything hardware can do, it can do faster than software."

On first impression, David Yen comes across as a modest, mild-mannered man who speaks easily about technology and less easily about himself. Those who know him well suggest there is a great deal of determination, humor, and passion just below the surface.

"He is very quiet, but don't mistake his lack of noise for lack of attention," says Sun CEO Scott McNealy. "He is listening and at the right moment, he will let you know what he is thinking and what the right course of action is."

McNealy credits Yen and his attention to detail for turning Sun's processor and network products group into "a much more predictable product-delivery machine." He also notes that the quality in Sun's silicon has improved dramatically under Yen's leadership.

"David doesn't bend to conventional wisdom -- he thinks it through," says Greg Papadopoulos, Sun's chief technology officer, who came to Sun because of an interview he had with the "amazingly passionate and wonderfully technical" Dr. Yen in 1994. "He seeks engineering truth. I think that's at the core of what makes him tick. At times when that puts him at odds with conventional wisdom, then he is contrarian. The baseline here is that he fundamentally respects great engineering. That's a pretty nice compass to have."

Sun now has an advantage, Yen says, because we were quick to recognize the coming "volume challenge" -- the fact that the Net will soon be populated by trillions of devices, including wireless handsets, RFID tags, and environmental sensors. We changed our approach to system design accordingly -- and acquired Afara Websystems in July of 2002 to accelerate the change.

"Without throughput computing, in the next two or three years, you could still get maybe two or three times faster performance, but that's not good enough," Yen says, "because the challenge is growing at a much faster pace."

With giant retailers such as Wal-Mart mandating the use of RFID by major suppliers, the next wave of network data is about to hit, even as the use of other wireless devices continues to pound data centers.

The conventional wisdom of simply lashing together "dirt cheap" commodity boards and boxes -- sold by weight in Asia -- isn't going to work. And the reason isn't so much technical as economic.

"Yeah, yeah, dirt cheap, all right. But the price still multiplies by how many boards you buy," Yen says. "If today's cost is X, which you barely can afford with your IT budget, and if you scale horizontally this way, you know the cost will increase proportionally.

"Two times, three times ... you might say, 'I can push harder, I can fight for more budget.' But if you talk about an increase of 1,000 times, nobody's IT budget is going to increase at that pace."

Yen is completely convinced that Sun has the right approach to throughput -- and so, apparently, is everyone else in the high-tech industry.

In fact, he says, the current situation breaks a longstanding pattern:

"Every time the company I work for goes out and says, 'We will do it this way,' usually the first thing competitors will do is say, 'That's wrong. That's useless. That's insignificant. My way is better.'

"We started talking about throughput computing at a time when Intel was out there with multiple executives in different forums, all talking about this commodity thing. They said, 'Processors are becoming a commodity and Sun doesn't know better and they will regret ... blah, blah, blah.'

"So we went out and talked about throughput computing. Notice, not a single competitor refutes Sun's argument. They know in their heart, once they cool down and think about it, this is real.

"This year, amazingly enough, you don't see the word commodity showing up in any Intel executive's speech. Not a single one. I've been monitoring very carefully. Instead, they all started talking about innovation. This year, finally, they concluded that it's not just a desirable thing but a necessary thing. This is why Intel executives, IBM executives, have all come out publicly, essentially endorsing the multicore/multithread approach.

"Now, how aggressively they want to pursue this sort of disruptive breakthrough, that's a separate issue. Different companies have different styles. Sun happens to have a culture where we are more aggressive in pursuing things we believe should be done."

What does that mean to customers? Sun's roadmap calls for 15 times the application throughput for network computing workloads by 2006.

Until now, other companies have had a significant advantage, just in terms of volume and cost, but these same competitors have serious disadvantages, too.

"If you are facing the need -- or opportunity -- for a paradigm shift, now the vertically integrated company has the advantage," Yen says.

"My processor guys talk to the Solaris engineers, so we work together, and this scenario is hard to do in the horizontally layered economy. The Microsoft, Intel, and HP engineers have to get together, yet everybody has their own agenda. It's difficult and takes much longer to accomplish anything.

"So any vertically integrated company in this situation can act, can execute. Yet in the whole industry right now there are only two such companies left: Sun and IBM. There is nobody else. Nobody else. This is even assuming IBM doesn't give up AIX."

Will Sun be able to make the most of this opportunity?

"It depends on the whole company -- from the processor group to the system group to the software group -- all executing together," Yen says. "So far it has been progressing very well. Engineers are breaking the boundaries of today's organization. None of the existing beliefs or practices fit exactly into this new scenario. So we are not trying to fit it into any of the existing modes."

And he's says he's not a contrarian.