x
Optical/IP

Juniper Mum on Core Router...

Juniper Networks Inc.’s (Nasdaq: JNPR) next-generation core router, code-named “Gibson,” is in trials with four carriers.

Two financial analysts, who didn't want to be named, have indicated that the product is in labs at Qwest Communications International Inc. (NYSE: Q) and NTT Verio, a subsidiary of Japan's NTT Corp.. Another source close to the company, who also requested anonymity, said that Cable & Wireless PLC (NYSE: CWP) and UUNet are also testing the product in their labs.

Industry scuttlebutt has it that Juniper will announce the new product in March, and that its official moniker will be the M320. It's unlikely to launch the device without being able to cite a customer. “I can’t comment specifically on unannounced products,” says Kevin Dillon, director of product marketing for Juniper. “But if you look at our history, what we have announced is real and in production. All of our announcements have customers.”

Sources say that the M320 uses the same silicon as its predecessor, the M160. If so, this means that it will inherit a packet ordering problem on the silicon used in Juniper's OC192 10-Gbit/s interfaces. The defect causes packets in a stream to arrive out of order in some circumstances.

Juniper maintains the packet reordering is not an issue in real networks. "It still appears at OC192, but only in limited applications," says Dillon. "In production networks it hasn't been a problem."

Two years ago Juniper announced its first core router upgrade, the M160, to its flagship router the M40 (see Juniper Sees Quadruple ). When the M160 was introduced, it was practically a no-brainer for customers looking to upgrade their networks. The M160 added four times the capacity of the M40, plus full hardware redundancy. Juniper easily signed up old customers and added new ones to its roster.

But the market has changed since 2000. Back then, service providers expected the M160 would be replaced with another upgrade within 18 to 24 months. Now, carriers are looking for a core router that will last them a minimum of three years, say experts.

“The market has definitely changed,” says David Newman, president of Network Test Inc., an independent testing company. “With capital spending so tight, carriers are trying to squeeze as much functionality out of their routers for as long as possible.”

A potential problem faced by Juniper is that the M320 might not have enough muscle to scale, observers say.

”Scaleability is the issue on everyone’s mind,” says Stephen Kamman, an equities analyst with CIBC World Markets. “But how scaleable is scaleable enough? That’s the big question.”

Those who have seen the new router report that it's an eight-slot chassis, which fits into half a standard seven-foot telco rack, with a total capacity of 320 Gbit/s. The plan is that four of these routers can be hooked together to create one logical router with a combined capacity of 1.28 Tbit/s.

But some wonder if this will be enough capacity, given that the Internet is growing 100 percent per year. Tier one carriers, which will be the first to adopt highly scaleable core routers, may need tens of terabits worth of capacity within the next couple of years.

Juniper-watchers say the company is developing a "super switch" fabric that will hook multiple M160 and M320 chassis together in order to deal with the problem. They expect it to be announced at the same time as, or just after, the M320. This architecture sounds similar to a system Cisco Systems Inc. (Nasdaq: CSCO) had under development in 2000, unofficially called the Terra Core. Cisco has never officially announced the Terra Core and has kept it out of its marketing discussions for over a year.

Cisco is also supposedly working on a next-generation scaleable router called “Q.” And Avici Systems Inc. (Nasdaq: AVCI; Frankfurt: BVC7) has been touting the scaleable story for over two years without much success (see Avici Intros Tiny TSR). Then there are the startups like Caspian Networks, Charlotte’s Web Networks Ltd., Hyperchip Inc., and Pluris Inc., which are all getting closer to introducing products into the market.

These newer players are taking a different approach to the problem. Unlike Juniper, which uses a centralized switch fabric in each device to process packets, these newer players have designed distributed routing systems. This means that some of the switch fabric in each device is resident on the line cards, and some is centralized as a means of interconnecting the cards. Software, too, is distributed, so the failure of one process does not take down the whole system but can be rebooted separately. This supposedly creates a "fault-tolerant" design, which is what many carriers say they want.

But some experts, including Network Test's Newman, say that it’s meaningless to debate architectures without products to compare.

“There are good arguments for each design,” he says. “I can think of ways to break either one. But what it will really come down to is implementation.”

Newman also points to the strength of Juniper’s customer base and says that many of these customers will choose Juniper because it is more familiar to them and cheaper than installing a new vendor. The cost of bringing in a new vendor can be anywhere between $12 and $20 million. These are costs associated with retraining engineers and integrating software into existing management platforms.

What’s more Juniper has already established itself as one of two main equipment providers in a highly competitive and specialized market, with a 25 percent to 30 percent share (see Router Numbers Support Cisco). (Cisco has the remaining 70 percent.) Tier-one carriers typically have about 40 points of presence (POPs), of which only 25 are likely to be mega-POPs where large core routers are needed. These carriers usually only buy two core routers per POP. At the end of the day, a core routing vendor is only looking at a maximum of 50 routers per customer over the next three to five years.

What’s more, the upgrade adoption rate is slow in this market. For example, one tier-one carrier has indicated that it only plans to upgrade four of its POPs this year. It will likely upgrade 10 more over the next year, and so on.

“This market requires a lot of support from equipment providers,” says Kamman. “And the trust involved is so great, I only see two or three players in the market, just because it is so highly specialized.”

— Marguerite Reardon, Senior Editor, Light Reading
http://www.lightreading.com
Page 1 / 2   >   >>
Belzebutt 12/4/2012 | 10:59:06 PM
re: Juniper Mum on Core Router... The numbers for the M160 are like this:
- 160 Gbps for the fabric (bidirectional?)
- 10 Gbps per slot
- 8 slots
So:
- 80 Gbps of actual interface capacity

For the M320:
- 320 Gbps for the fabric
- 8 slots
So...
Does Gibson only support 20 Gig or interface traffic per slot?
If Gibson support 40-Gig slots, i.e. each slot has 4 times the capacity of an M160 slot, then why is it called the M320?
Did they make the box only capable of operating at half line-rate, or did they do the math differently?
Should it really be called the M640, is it really only twice the capacity of an M160, or do the old boxes do "fuzzy math"?
Belzebutt 12/4/2012 | 10:59:05 PM
re: Juniper Mum on Core Router... Yes but the M160 is half a rack too. How come the M320 supports 4x the slot capacity if it only has 2x the switch fabric?
myresearch 12/4/2012 | 10:59:05 PM
re: Juniper Mum on Core Router... It's half of a rack so it is 4x10G per slot and 8 slots.

MR
metroshark 12/4/2012 | 10:59:04 PM
re: Juniper Mum on Core Router... Let's see:

- The packet processor on M40 can handle 8 OC48 ports, or 20Gb/s of full-duplex traffic
- M160 uses 4 of these processors in parallel to handle 8 OC192 ports, or 80Gb/s of full-duplex traffic. It is still 20Gb/s per packet processor though.
- Both M40 and M160 use a centralized packet forwarding architecture.

Based on this, my best guess is that for M320, Juniper changed from a centralized architecture to a distributed architecture. All they have to do is use one of the M40/M160 packet processors on each line card and connect these line cards with a centralized switch fabric. They could easily build a 8x20G or 16x20G system like this using most of their existing forwarding chips and some new queuing and switch fabric chips.
wdog 12/4/2012 | 10:59:03 PM
re: Juniper Mum on Core Router... Changing architectures is never simple. It isn't about what hardware can be re-used, it's about the software. If the software was architected for a centralized system it doesn't magically work in a distributed system. Much of the code, much of JUNOS, has to be re-written to support a new architecture. Whatever the M320 turns out to be, odds are high that Juniper will have to make trade-offs in schedule versus a new architecture. My bet would be they go for a better schedule and do what they can with exiting JUNOS.
skeptic 12/4/2012 | 10:59:03 PM
re: Juniper Mum on Core Router... These newer players are taking a different approach to the problem. Unlike Juniper, which uses a centralized switch fabric in each device to process packets, these newer players have designed distributed routing systems.
-------------

I would say that "some" of the newer players
are doing architectures like this, but certainly
not all. You sound like you have been talking
to Pluris a whole lot.

------------
This means that some of the switch fabric in each device is resident on the line cards, and some is centralized as a means of interconnecting the cards. Software, too, is distributed, so the failure of one process does not take down the whole system but can be rebooted separately. This supposedly creates a "fault-tolerant" design, which is what many carriers say they want.
----------------------
This is wrong. The failures where one process
takes down another are not really solved in a
distributed system. All a distributed system
will protect against (without further work) is
failure of hardware components where the software
is running.

flanker 12/4/2012 | 10:58:53 PM
re: Juniper Mum on Core Router... The numbers for the M160 are like this:
- 160 Gbps for the fabric (bidirectional?)
- 10 Gbps per slot
- 8 slots
So:
- 80 Gbps of actual interface capacity


According to the spec sheet its full duplex throughput on the M160 per PIC. As you said 10 GBPs (duplex) per slot. I would agree that that is only 80Gbs of unidirectional capacity.

According to myresearch, the slot capacity has increased four fold (4 x 10) per slot. If the slot capacity has increased 4 fold, then yeah, sounds like it should be called an M640 to be consistent.












indianajones 12/4/2012 | 10:58:09 PM
re: Juniper Mum on Core Router... I agree with the poster who surmised that the Internet Processor II ASIC ism oved to the line cards. That makes sense since the IP-II was capable of forwarding at 40G and if they are really building a 40G card that fits in well. I also read reports that the first line card that they will support is the 1-port OC-192 card which means that when the product comes out it will support only 8 OC-192 ports like the M160. Really confused!!

In any case, I will start believing when they deliver. Remember, even Tony Li did not think that the ASIC engineers at Juniper were competent. How are they going to pack 32 10G in a half a bay?
Lopez 12/4/2012 | 10:58:08 PM
re: Juniper Mum on Core Router... I agree with the poster who surmised that the Internet Processor II ASIC ism oved to the line cards. That makes sense since the IP-II was capable of forwarding at 40G and if they are really building a 40G card that fits in well. I also read reports that the first line card that they will support is the 1-port OC-192 card which means that when the product comes out it will support only 8 OC-192 ports like the M160. Really confused!!

Perhaps they are only supporting 1xOC192 because there is no market for anything else?

I expect that Gibson is so slow in coming because Juniper has moved resources off it and onto projects that are going to make the company money. You know all those "edge PICS" that were announced right after this article, those are the things that are making Juniper money.

There is no market for Gibson in the near future.
mrcasual 12/4/2012 | 10:58:07 PM
re: Juniper Mum on Core Router... Most of the discussion has been around whether Gibson CAN support 1x40G. Personally I think the more interesting question is WHEN should it support 1 x 40G.

40G slot capability is generally a good thing because it should let you deploy multiport interfaces (192c, 48c, etc) at a lower cost per port.

However, a single 40G port is a different beastie all together and brings a lot of questions such as:

Is there a pipe that really needs it?
Is there transport infratructure in the ground/pipe capable of supporting it between two sites?
Are there optics and framers available?
Is anyone willing to pay the certainly high price in this market?

Can anyone offer any insight?
Page 1 / 2   >   >>
HOME
Sign In
SEARCH
CLOSE
MORE
CLOSE