Packet inspection/traffic management

Anagran Promises TLC for TCP

It's not Caspian. That's one of the main things Lawrence Roberts wants everyone to know about his new company, Anagran Inc.

But there's plenty more that Roberts, founder of now-defunct IP equipment vendor Caspian Networks, is keen to say about Anagran, which today unveils its maiden product, the FR-1000 Intelligent Flow Router.

The FR-1000, says Roberts, is an edge router that can control traffic more efficiently than deep packet inspection (DPI) products, allowing TCP networks to run at 95 percent of capacity instead of the usual 30 percent. It's a 1U box with 48 Gbit/s of capacity (not double-counting inputs and outputs), and has its roots in the Flow-Based Networking that was also the basis of Caspian Networks's router.

But Roberts left Caspian more than three years ago, convinced the company wouldn't fly. And he was right; Caspian folded last year. (See Caspian Closes Its Doors.)

Roberts is known for a lot more than Caspian. He's part of that elite group that helped build the Internet in the first place. (See Dr. Lawrence Roberts.) And, with Robertsian flair, he's willing to proclaim that Anagran could "shift the Internet to a fundamentally new technology."

But he's worried that people will dismiss Anagran as a Caspian clone.

"Caspian came at a time when it looked like it should be a core box," Roberts says. "Their margin was too low, even though the product worked and everybody was happy with it. I had quit halfway through because I saw the margins would be zip, and I thought there were a lot of other things we could do with the technology."

Compared with Caspian's costs, Anagran comes cheap. The company first came to light in 2004 as Roberts collected his early funding. Three years later, investors have put just $28 million into the 75-employee company. Caspian's third round alone was worth three times that. (See Supercomm Snippets: The Sequel, Anagran Raises $8M, and Anagran Attracts $12M.)

Roberts was able to keep development costs low by relying on off-the-shelf chips, something startups couldn't necessarily do circa 1999. The heart of the FR-1000 is an FPGA programmed with Anagran's algorithms, alongside what Roberts describes as an inexpensive routing chip. All told, the box lists for $70,000.

Anagran isn't just an exercise in pinching pennies, though. Roberts says he's spent the last few years expanding on what he started with Caspian, coming up with a more powerful way to use flow-based networking.

So, what is the FR-1000?

Inside the box
Anagran's router was conceived as an add-on to live networks, something that would be deployed as an aggregation box at the network edge.

Like Caspian, Anagran uses flow-based routing, where the router processes traffic in terms of flows and not just by individual packets. But this time, Anagran adds the wrinkle of being able to control the bandwidth used by each flow.

The goal is to avoid the high packet loss that comes when the network gets overloaded. Put differently, Anagran claims to turn TCP into a better juggler, letting the network maintain high throughput on a higher number of flows.

When a normal router reaches overcapacity, it starts discarding packets, often in a round-robin fashion, sometimes affecting traffic flows that aren't contributing to the problem. "The result is, everybody slows down at once. Then, they all start speeding up at once," Roberts says. To avoid this oscillatory behavior and the congestion that comes with it, carriers run their routers at around 30 percent of capacity, to leave enough headroom for surges in traffic, according to Roberts.

To Page 2

1 of 2
Next Page
Page 1 / 2   >   >>
aswath 12/5/2012 | 3:04:25 PM
re: Anagran Promises TLC for TCP A snip from the article: "Video gets a higher rate than voice, for instance. Then, if a particular flow approaches the assigned maximum, the router drops one packet from that flow, slowing it down without affecting any others."
Pete Baldwin 12/5/2012 | 3:04:25 PM
re: Anagran Promises TLC for TCP Interesting quote from 2001:

"Larry Roberts is no Dave Huber and Caspian is no Corvis." (That was actually meant to be an insult.)

Pete Baldwin 12/5/2012 | 3:04:25 PM
re: Anagran Promises TLC for TCP The idea is certainly intriguing. I'll be interested to see how much uptake it gets.

Larry Roberts says he envisioned Anagran as an add-on to the edge of installed networks, which makes sense. But in Asia, apparently they've come up with a greenfield *core* application.

The idea is to build an optical core with FR-1000s around the perimeter. So now, you've got no routing and no MPLS in the core, just the FR-1000 directing traffic and pointing it to the other side.

I may be oversimplifying that, and Anagran didn't share details like how big a core network we're talking about here. Makes for an interesting architecture, though.
Occums razor 12/5/2012 | 3:04:25 PM
re: Anagran Promises TLC for TCP LR strikes again!! Go with the flow man.

I can't believe anyone gave this guy more money to loose. I am genuinely shocked.

He's right when he says that the Internet business model is broken, so why then does he keep building equipment that ISP's don't need and can't afford to buy becuase their business model sucks! From his last venture he said the margins were poor - no sh*t the margins are poor becuase the business model sucks, the business model sucks, the business model sucks, the business model sucks - get it.

Why spend more VC money building boxes that ISP's just can't afford to buy becuase their business model sucks? I wonder how long it will take Dr Larry to figure out that his margins suck yet again, duh?

The point is that ISP's will always find it easer and cheaper and less hassle just to throw bandwidth at the problem, this is the way the Internet is built, its the way it has always been built, it's called occums razor, usually the simplist solution is the best solution. The simplist enginering solution is always to add more bandwidth, with enough bandwidth all applications run just fine, QoS is not needed, with enough bandwidth everything runs just fine, like it does in the Internet right now.

Anagran may be able to improve link utilisation under some cicumstances but I sinceley doubt that any improvement is likely to cost justify ISP's adding more boxes to their networks. With more cost, more complexity, more things to break, at the end of the day ISP's are hard pressed financially, when it comes right down to it I think they'll find 20 things much more pressing to spend money on than this.

Just my 2 cents :-) If anyone can tell me different I'm all ears
fgoldstein 12/5/2012 | 3:04:24 PM
re: Anagran Promises TLC for TCP How will flow-based routing mix with others on the same network?

Oscillatory "slow start" packet behavior is somewhat inefficient, but if everybody follows it, it's also rather "fair" -- resources get allocated a certain way, and it's mathematically stable. A network where all traffic follows the TCP rule is never going to have congestion collapse. It can also run at a fairly high occupancy (well over 30%), but of course it will have packet loss, precisely because that is normal.

If you have loss-sensitive applications (streaming voice, for instance), then the overall usage level (%) has to go down, or there has to be some mechanism to separate or prioritize traffic. This is the subject of much debate.

If you have a mix of slow-start and noncompliant apps, though, then the compliant slow-start TCP apps will likely be the losers. Streaming and more clever rate-adaptive schemes will both take a larger share, leaving TCP with less. This "selfish optimization" (using Raj Jain's terms -- he's Da Man for this type of theoretic stuff) typically lowers overall network throughput.

That's why "IP under alles" is such a tricky business. IP wasn't meant for it. Flow-based systems of some sort do make sense, but they have to be seen in context of the entire network that they're used in.
^Eagle^ 12/5/2012 | 3:04:24 PM
re: Anagran Promises TLC for TCP Throwing bandwidth at the problem sounds simple dear mr. occam.

but it is only simple on paper.

Bandwidth is expensive, especially bandwidth that runs at high speed (OC192, OC768). Even OC 48 is very high priced.

Also, be aware the market for this box is not for poor struggling ISP's. The market is for CARRIERs! The carriers who have to guarantee bandwidth performance on the links that the carriers sell to ISP's.

While the independent ISP model is broken, the carrier model is not. Carriers are making money and are facing increasing capacity limitations in the core. Remember, most of the "overinstalled" capacity that was put in place during the late 90's and into 2001 is now used up and gone.

better link utilization for carrier IP transport networks would be a BIG savings to the carrier who must over provision all those IP links from Cisco and Juniper routers just to deal with the poor efficiency of standard IP transport methods.


p.s., I don't work for Anagram. Did not work for Caspian. Do not know and have never met Larry. just my own humble opinion
schlettie 12/5/2012 | 3:04:24 PM
re: Anagran Promises TLC for TCP Core router margins are too low? Since when!
paolo.franzoi 12/5/2012 | 3:04:24 PM
re: Anagran Promises TLC for TCP
Just as a correction: It is generally noted as Occam's Razor. Interestingly enough, it was created by William Ockham.


Pete Baldwin 12/5/2012 | 3:04:24 PM
re: Anagran Promises TLC for TCP Good points. I don't think margins will be as tough as Caspian's (that company had more than 300 people!) but Anagran will have to compete with a lot of conventional thinking that, as Vorhaus put it, isn't considered broken yet.

I can't resist pointing out, though -- it's spelled Occam's Razor (or Ockham's), and it applies to the explanation of past phenomena, rather than the best way to plan the future.
mr zippy 12/5/2012 | 3:04:23 PM
re: Anagran Promises TLC for TCP by setting up many concurrent flows to different destinations? The flows don't have to be big or long lived, it's the flow setup processing and memory allocation to follow them that I think would be the weakness.

This is the reason why routers have moved away from route caching towards full RIB->FIBs e.g. CEF. From what I understand, that happened in the network core because lots of flows, and therefore diverse routes to cache, is the nature of traffic that core routers see. At the edge, the number of flows would be much lower, however, you're then vulnerable to kids running random flow setup software for fun, and that software is quite easy to build or find. One of the reasons it's easy to find is that it is used to see how well a router can handle random flows! I understand that Linux still uses route caching (for a end host it still makes some sense), and I've seen mention on the Linux network development mailing list of random traffic destination testing tools to evaluate it's performance.
Page 1 / 2   >   >>
Sign In