Clean Slate

New Networking Stack Does What Others Can't

Story by Al Riske. Photography by Howard Friedenberg.

10.July.07- Sun has always believed that computers should all communicate as one, no matter who built them. So that people can all communicate. Can all do business. Can all participate, share, and collaborate.

The company became famous for declaring, long before it was evident to others, that "the network is the computer."

That much is well known.

But as Sunay Tripathi puts it: "You can never say, 'Okay, I'm the leader now. I'm done.'"

What's less well known is that this 10-year Sun veteran has been a key player in helping Sun to live up to its slogan and once again forge ahead of the pack.

He's also the man with a "slightly wicked plan" to help the company recapture a market it walked away from 15 years ago.

After working as a researcher at both the Indian Institute of Technology, Delhi, and Stanford University, Tripathi joined Sun in 1997.

"I came here because I knew people who worked in Solaris networking. Big names. The leaders in this space," Tripathi says. "Then around 1999 or 2000 a lot of these people left for whatever reason and moved to different groups or different companies. Suddenly I realized, 'Uh-oh, there's no one behind me.'"

Which made his next decision more difficult and, in a way, easier.

"I kept saying, 'Why the hell are we doing things this way?' We had a stack that was 20 years old, designed by AT&T for a totally different model -- the Web didn't exist at that time -- and we were still trying to patch it and fix it," he says.

This is where maybe I was just too naïve or stupid, but I didn't carry a lot of the baggage that other people carried. I took the decision that, 'Sorry, we're just going to rip this out and rebuild it from zero.'"

It was a bold decision for someone so new to the company, so young, and so junior, but circumstances worked in his favor: The economy was in a tailspin and Sun had its back against the wall.

Perfect.

To make matters worse, competitors started referring to Solaris as "Slow-aris" and the label was beginning to stick.

"Jonathan Schwartz sent an email to me saying, 'Can we get rid of this Slow-aris label somehow?' I said, 'Yeah, sure, we have a plan to do that,'" Tripathi recalls.

Here the Distinguished Engineer uses a broad-brush analogy to sum up a complex undertaking: "A 1965 Ford Mustang is a great car. But if you want to go to the moon, I don't think that's the right vehicle. You'd better design a space craft."

So Tripathi pushed ahead, with support from top engineers such as Tim Marsland and Erik Nordmark, on a clean-slate design. The turning point came two years later with the release of the Solaris 10 operating system.

"We delivered the first part of the strategy, Project FireEngine, which dealt with TCP/IP performance, and we got rid of the Slow-aris label as part of that," Tripathi says. "At that point people realized, 'Okay, this is going to work. The strategy has traction.' From then, things became a lot easier -- and I wasn't that immature anymore. I was a lot more ... "

Politically savvy?

Tripathi laughs. "A couple of my mentors reminded me that, 'Sunay, you are a friendly person. Why do you get so intense when you are talking to people about technology? Yes, enthusiasms is right, but can you tone down your intensity and keep your friendly face there?' People are more open to change if it's coming from a friendly place, right?"

All of which simply sets the stage for the real story: another important leap forward in the Solaris networking stack, Crossbow.

"With Crossbow, we can do network virtualization and resource partitioning in a very simple manner, without any performance penalty. Zero. That's a big change," Tripathi says.

It's huge, actually, because people expect to take a performance hit with virtualization.

Then comes the realization that, thanks to Crossbow, Solaris software now provides the key features of a network router. A very fast router.

But the real attention-getter is this: Let's say your firewall, load-balancer, and DNS server all come from different vendors and you're spending a lot of money for them. "Now ," Tripathi says, "you can actually get an open-source version, on Solaris, combine them on one box, and partition the box so each gets the resources it needs." (See Project Virtual Network Machines.)

The potential cost savings are significant, to say the least.

"The idea is to do to the networking vendors the same thing that Linux did to us," Tripathi says. "Solaris is a general-purpose OS. It's free. It's runs on cheap hardware. Why would you want to pay the networking vendors a boatload of cash?"

Seven years ago, Tripathi laid the groundwork. He worked closely with the leading makers of network interface cards, or NICs, asking for design changes -- basically, a different way of laying things out -- that would cost them nothing.

Broadcom, Neterion, and Intel all signed up. And we added the functionality to our own NICs as well.

The payoff? Tripathi compares the new design to a freeway with very sophisticated metering lights. Metering lights that direct traffic into specific lanes for specific purposes.

"Compare that to a system where cars are already on the freeway and you have to sort them out into different lanes. That's a big problem. But if you get them into the right lane when they enter the freeway -- that's what we are doing. So there's no performance hit," he explains. "We enforce a lane discipline as soon as the packet enters the NIC or leaves the NIC. So there's no extra code. When a packet comes in we're deciding where it goes and, based on that, everything flies."

The same approach boosts performance of our multicore UltraSPARC T1 ("Niagara") processors.

"When a packet comes in, we see which core is less busy and we just guide the packet to that core. In other words, we have the ability to parallelize network traffic on the cores of Niagara," he says. "That's a big win for us."

Tripathi notes that Sun's early systems were often used as routers but the company considered that a niche market at the time and chose to focus on workstations and servers. Sun has since tried to re-enter the market with various products and acquisitions, but has found it difficult to gain traction.

"The strategy is not to go head to head with Cisco but to target alternative markets. Cisco doesn't control the telco market or the NEP [network equipment provider] market," Tripathi says.

"Telcos and NEPs don't want to depend on Cisco. They don't want to depend on anyone. They want to be independent of anyone pulling their strings. So they are all considering, you know, can we use Sun and Solaris in this space and actually be independent of some of the big vendors and have an open-source solution?"

Sun can also target independent software vendors who could compete more effectively running on top of Solaris.

"It's a slightly wicked strategy," Tripathi says. "Let the others do the fighting for you. You be the arms supplier."

The strategy appears to be gaining traction, too.