networkZONE Products for the week of April 26, 2004
Greenfield Networks Says . . .
Stalking Cisco: Greenfield Networks' Full-Featured
Modular Switch Chip Set Targets Cat-3500/Cat-6500-Class Products
Industry's First Ethernet Switch Silicon Includes Full Support
For IPv6, MPLS and Layer 2/3 VPNs; Addresses Broad Range of Enterprise and
Metro Markets
Greenfield Networks announced the availability of the industry's most advanced Ethernet switching silicon solution. Featuring full IPv6 routing in hardware, Greenfield's breakthrough Packetry architecture supports high-density Gigabit performance with the breadth of features and flexibility to scale from the edge of the enterprise to metro area networks.
Greenfield Networks is funded by preeminent venture firms Sequoia Capital, Global Catalyst Partners and Walden International. The founding team of industry experts has a proven track record of delivering innovation, most recently at Cisco Systems, where they worked together on the development of many Catalyst switching products.
"The enterprise networking market is moving rapidly towards Gigabit Ethernet on the desktop, with major industry initiatives for expanding voice and video capabilities over IP. Likewise the worldwide Metro Ethernet market is gaining traction, and absolutely requires support for MPLS and Layer 2/3 VPNs," said Bob Wheeler, senior analyst at The Linley Group. "IPv6 is now a mandated requirement for networks deployed in many countries and within many governments, including the US federal government. For high-end Ethernet systems, Greenfield is very well positioned to take advantage of the emerging market opportunities."
The Packetry Concept
Greenfield's Packetry family provides a unique architecture for scaling
performance from edge to core applications. It is the only merchant silicon
in the industry to offer the combination of low-cost Gigabit Ethernet connectivity
and rich features, such as full IPv6 routing, MPLS switching, scalable packet
buffering, and scalable MAC, route and classification tables.
"The demand for full IPv6 support, including routing, QoS, access control lists, and IPv4 to IPv6 translation, is an essential capability for high performance Ethernet systems," said Gary Smerdon, president and CEO of Greenfield. "We are the first silicon provider to deliver a world-class 'no excuses' solution."
The Packetry family offers the most feature-rich Ethernet switching solution in the industry. Packetry delivers robust Layer 2/3/4 packet processing capabilities, including IPv6 routing in hardware and extensive Quality of Service (QoS). Packetry's classification and traffic management functions enable voice over IP and other converged applications that require flow-based priority and bandwidth distribution across the network.
Enabling the next-generation metro edge deployments, the Packetry family supports extensive service provider features including MPLS switching, stacked VLANs, Layer 2 VPN (martini and VPLS), Layer 3 VPN (RFC 2547bis) and IP tunneling. The Packetry family has the architectural flexibility to address a wide range of platforms, from fixed configuration "pizza boxes" to centralized edge chassis and high-end distributed core chassis systems. This flexibility, coupled with the extensive feature set, allows the Packetry devices to address a broad range of markets, from the desktop edge of the enterprise to the metro area network.
Greenfield has already shipped product to five countries, and has seen its Packetry family designed into products in all of its key market applications. The worldwide system opportunity for the Greenfield target markets is estimated to be more than $10 billion.
The Packetry Difference
Greenfield Networks has approached product development with end systems
integration in view, putting tremendous up-front focus on architectural
emulation and software development in addition to chip design. Greenfield
provides a complete system level solution to reduce development costs and
accelerate time-to-market, including reference design systems, production
ready API software, and interfaces to 3rd party control plane software.
Designed from the ground up, the Packetry API is completely isolated from hardware details. The API accelerates time to market by freeing customers from needing to understand register and data structure details. The API provides a complete functional abstraction of all Layer 2/3/4 functions, and maintains hardware, protocol stack and OS independence.
"I am pleased that we have been able to deliver all out advanced functionality with first revision silicon," said Harish Devanagondi, co-founder and CTO of Greenfield. "I attribute this to the strength of the Greenfield team, and the development methodology we put in place, including live network emulation and hardware-software co-development."
The Packetry Products
There are three initial members of the Packetry family: the G525 Packet
Engine, the G750 Shared Memory Fabric, and the G120 Intelligent Multiplexer.
All devices are available now.
G525 Packet Engine
The G525 is a packet processing engine capable of wire speed operation for
16GE ports, or one 10GE/OC-192 port. The G525 packet engine provides packet
processing for Layer 2/3/4, IPv6 & MPLS, and also supports advanced
services such as QoS classification, packet filtering and Layer 2/3 VPNs.
The G525 interfaces to an external TCAM to provide scalable MAC, IPv4/v6
route tables and classification/filtering tables.
G750 Shared Memory Switch Fabric
The G750 is a single-chip 32G shared memory switch fabric/traffic manager
that enables a variety of buffering and queuing functions including fair
queuing, scheduling, traffic shaping, congestion control, multicast and
resource management. The G750 provides scalable external packet buffering
up to 256MB. The G750 may be used in fixed configuration switches or mid-range
chassis systems as the central switch fabric of the system, or in a large
distributed chassis as a traffic manager on a linecard.
G120 Intelligent Mux
The G120 is an intelligent multiplexer device used to develop high-density
and ultra low-cost Gigabit Ethernet linecards for edge chassis systems.
The G120 supports advanced multi-tuple classification, policing, and class-based
queuing to guarantee bandwidth for high priority traffic. The G120 can interface
with up to 128MB of external packet buffer memory to absorb short-term congestion
in the network.
analogZONE Says . . .
Dear Readers,
Please accept my apologies, in advance, for what may be the longest review I've written to date. If I could have come up with a shorter way of discussing Greenfield's products, what they do, and what their potential market impact was, I'd have done it. Unfortunately, it seems that as the complexity of the products I cover increases, the level of detail in manufacturer's press releases decreases. This puts me in the difficult position of having to both explain many important details left out of the release as well as provide some analysis of the chip's place in the market. Hopefully, you'll find the rather long-winded report below of some use. If not, please let me know.
Until now, the high-end networking market has been relatively immune from the commoditization that has seen unmanaged 10/100 SoHo and workgroup equipment prices plunge below $20/port. But the high levels of custom functionality, deep intelligence, and raw throughput that served as steep barriers to entry for merchant chip makers and ODMs may finally be breached with the introduction of Greenfield Network's IPv6-capable, LECE-class (Large Enterprise, Carrier Edge) chip set.
Rather than try to compete with the likes of Broadcom, Marvell, and Vitesse for sockets in the fixed-configuration commodity L2/L3 switches and "pizza boxes" that lurk in the wiring closets of so many businesses, they are making an assault on a domain that's been historically dominated by ASIC-heavy Cat-3500/Cat-6500-class modular boxes.
Greenfield is targeting the bigger, more profitable opportunities with a family of IPv6-capable Ethernet switching devices that will enable equipment makers to address a broad range of medium-to-high-end applications.
Designers will find themselves with a new set of building blocks that allows them to use the same chip set to produce several classes of products. This includes high-capacity fixed-configuration aggregation boxes (used in WAN edge, or enterprise L2 aggregation). Further up the food chain, the same chips will power flexible midrange modular edge chassis (Greenfield calls them "desktop chassis") that are used in enterprise environments as a more flexible alternative to stacking switches. At the upper end of the range manufacturers will have the freedom to offer a general-purpose chassis that can support advanced features (firewalls, load balancing VPNs, VoIP gateways), on a per blade basis at or below the per-port cost of a fixed-function box.
Design Drivers
To understand how Greenfield intends to make good on its intentions, you
have to take a look at the factors that they considered when architecting
their chips. On the top of the chip, designers' must-have features list
were native IPv6 processing, 10-Gbit/s uplinks, low-cost, high-density 1-Gbit/s
desktop connections, plus support for voice and video as well as metro services
& wireless. The other driver for the design was the need for modularity
that would allow designers to create application-specific boxes without
massive re-engineering or the use of costly ASICs.
Taking the requirements from the top, IPv6 capability is going to become a must-have feature for any equipment hanging near the WAN edge within the next few years. It's already required throughout China and Asia thanks to the explosion of IP addresses in the region (and top-down governments that can make national IT policy). We're seeing a similar, but less urgent push to upgrade our North American networks as new DoD requirements are coming into effect and larger numbers of personal electronic devices hit the Internet and start requiring their own IP addresses.
While not essential in low-end commodity boxes, native IPv6 capability will soon be an important feature for future-proofing higher-ticket capital equipment with a minimum service lifetime of 5 - 7 years. In much the same way, mid-range switch silicon will also be required to efficiently handle MPLS traffic to support the many advanced features that modern networks are expected to deliver.
I agree with Greenfield's assertion that modularity could become an important feature in many places where fixed-function boxes have dominated until now. They rightly observe that modularity is one of the best ways for manufacturers to meet the extremely varied demands the LAN is beginning to play in the enterprise and SMB market.
While many networks are still simply shoving data around between PCs and Servers, there is a growing trend to add VoIP traffic and IP-PBX functionality to the mix. MIS managers are beginning to demand boxes that can be easily upgraded to support these features via an add-in card or module. The increasing popularity of devices like blade servers, load balancers, and firewalls all demand that these functions be easily and economically provided as upgrades to existing customer equipment rather than more expensive separate units.
Likewise, the steadily-increasing use of Power-over Ethernet (PoE) to drive Wi-Fi access points, web cameras, and IP phones means that today's boxes must easily upgrade to PoE on an as-needed basis.
Tough Requirements, Tough Silicon
These tough requirements gave rise to a modular chip set that uses a centralized
switching scheme for most applications, but can be scaled to fractional-Terabit
size using commercial crossbar fabrics to form larger meshes. The basic
set of building blocks consists of a packet processing engine (the G5xx
series), shared memory fabric/traffic manager supervisor (the G7xx family),
and an intelligent mux which provides efficient aggregation of multiple
Gigabit Ethernet interfaces (the G1xx).
Greenfield's architecture calls for the system's switching intelligence to reside centrally in the G7/G5 elements, while the ports (usually mounted on blades or modules) simply handle line interfaces. Since the chip set was designed with an eye towards both LAN and infrastructure applications, the control intelligence can be either a single-point, or a 2-for-1 redundant configuration.
The G525 Packet Engine
The first incarnation of Greenfield's packet engine is the G525. It does
all the classification and routing at wire speed for 16 10/100/100Base Ethernet
MACs or a 10-Gbit/s SPI 4.2 connection before passing it on to the G750
switch fabric. The SPI 4.2 line side interface can hook up to a10GE MAC,
a SONET framer or even an NPU. This, plus the ability to terminate both
PPP and HDLC connections, allows the G525 to support both Packet and SONET
data with equal ease and no-exception processing. Among other things, this
means that you get a native OC-192 WAN connection at LAN prices.
The controller-to-fabric connection is handled through a 550-MHz DDR SPI 4.1 interface that supports well in excess of the 16 Gbit/s coming from a fully-loaded set of inputs. This allows the switch to run at full capacity with room for the overhead caused by added tagging headers that are used internally.
One of the G525's most striking features is its support native IPv6 and MPLS traffic without control plane intervention -- something usually found only in home-grown ASICs used by "heavy iron." You also get many other nifty features like label edge routing (popping up, encapsulating, and swapping labels in VPN and MPLS applications).
Greenfield also differentiates its device from commodity switches by supporting the full range of features found in enterprise and metro access systems. The "laundry list" of the protocols and standards it supports is far too long to publish in full here, but it includes:
The G525 packet engine also packs a bunch of intelligence to handle the advanced functions that can greatly improve QoS, enable advanced policy enforcement, provide wire-speed security, and make more efficient use of the system's available bandwidth. Among the top features in the hit parade are:
The G525 uses external RLDRAM memory to store its classification tables, MAC addresses, and other lists that are accessed using a pointer generated by an attached TCAM. In fact, the TCAM handles all lookup tasks for classification, filtering and access control functions. While slightly more costly, using commercial TCAMs gives you both speed and the ability to scale your tables to any size. Depending on your application, you can attach devices ranging from 2 to 18 Mbyte, and even cascade them if needed. Being able to perform deep wire-speed searches on large lists is a big plus in wireless data, or other public access applications at the WAN edge which usually involve tens or hundreds of thousands of users and huge route tables.
Greenfield is the first commercial switch maker to employ commercial TCAMs (from Netlogic or IDT) for managing its tables and lookups. While network processors have used these devices for some time, most commercial switch silicon (other than SwitchCore, which uses a proprietary device) relies on internal lookup structures or algorithmic tricks to approximate TCAM functions using fast RAM.
The G750 Switch Fabric
The G750 shared memory fabric family hooks up to one two G525 packet processors
to support 16 or 36 Gbit/s worth of full duplex, wire-speed switching capability.
The device uses tags generated by the G525 to perform both buffer and queue
management. Its shared memory architecture uses up to 256 Mbyte of 300 MHz
DDR RLD RAM, allowing support and buffering for lots (and lots) of queues.
The deep, flexibly-segmented memory allows you to create deep buffers that
minimize TCP-imposed slow-downs that are caused by dropped packets. Its
deep buffers are also good in VoIP systems since they allow low-priority
non-voice packets to dwell longer before being dropped.
The G750's centralized queuing scheme uses tagging information from the controller to perform advanced traffic management for up to 32 Gbit/s ports in a pizza box or edge chassis system. You can expand the port count beyond this using either third-party crossbar silicon or, if bandwidth is not an issue, the G125 multiplexer chips (more on this shortly).
There are also 32 dedicated queues which are used to provide the host CPU with access to any data stream running on the fabric. This hardware stream processing helps the host prioritize tasks without any processing overhead, and also reduces its workload when coping with DOS attacks and other high-stress situations.
To accommodate the stiff requirements of enterprise and WAN applications, Greenfield has incorporated a variety of intelligent mechanisms that support quality-aware, class-based, QoS and queuing using the 802.1p/IP DSCP/IP TOS/MPLS EXP. Some of the features that help the G750 handle large numbers of mixed-priority streams include:
Finally, it's important to note that any and all of these features can be invoked simultaneously without affecting throughput. This is not so for many of the lower-end switch chips currently on the market which have various degrees of "feature interaction". As an example, they have demonstrated how the controller can perform 3 MPLS label pops, an IPv6 ACL operation & classification, an Ipv6 forwarding lookup, and then pass the packet to its egress port where an IPv6 ACL lookup is also performed -- all at wire speed on all ports.
The G120 Intelligent Port Multiplexer
While not as glamorous as its companion parts, the G120 port multiplexer
is a sophisticated and efficient way to expand the port count of a system
at a rock-bottom price. The device can be programmed to aggregate Gigabit
Ethernet traffic in either an 8:1, 8:2, 4:1 multiplexer configuration. It
has eight integrated RGMII MACs on the line-side interfaces, and two more
for passing the data on to the G525.
The G120 takes advantage of the fact that most applications use only a fraction of a Gigabit connection most of the time to provide intelligent stat-muxing that makes efficient use of the pipelines without significantly impacting service. It does this using a combination of local classification, policing, and class-based queuing to make sure the high-priority traffic makes it through with a minimum of delay.
The device uses a large chunk of PC RAM for the buffering that supports efficient and intelligent over-subscription. The programmable flow control mechanism can support up to eight classes of service per port while enforcing a had guarantee of up to 256 Mbit/s worth of bandwidth for each one under any and all load conditions. All of these functions, plus optional flow based load balancing across uplinks and VLAN STP state handling are performed locally, offloading them from the switch itself.
This intelligent over-subscription is excellent for edge aggregation applications as well as other applications that must make the most efficient use of available bandwidth, or strike a balance between cost and performance.
Putting It All Together
Greenfield's centralized architecture delivers the functionality of Cat-3500/Cat-6500
class modular boxes at not much more than the cost of a stackable pizza
box. The chip set is equally at home in low-cost, full-featured, fixed-feature
mid-range managed chassis (pizza boxes), and configurable (blade-based)
edge equipment that can retail for under $100/port. It also offers designers
the option of creating "modular pizza boxes" such as a 16-port
GigE chassis with a limited bank of swappable interface cards. And since
the switch can split its bandwidth between IP and SONET seamlessly, (using
PoS over PPP) a manufacturer can even offer an OC-192 metro interface at
a price point unheard of today. I expect that thanks to Greenfield, Metro
Ethernet equipment is about to get a whole bunch cheaper.
The chip set also allows you to build larger distributed switches for core applications (up to 640 Gbits/s capacity) using 3rd-party merchant switch fabrics. In this mode, the 525 & 750 then serve as intelligent line cards and talk to the switch fabric via the 750's SPI 4.2 interface. As mentioned earlier, the controller/fabric interface can handle well over the nominal 10 Gbit/s (16 Gbit/s, and a little more). This is essential for wire-speed operation at 10 Gbit/s because of the extra freight imposed by the tagging, encapsulation and other padding used by crossbar mechanisms.
As mentioned earlier, Greenfield has stressed that that the chip set can support virtually all of these advanced features simultaneously -- something that other architectures that rely on control plane intervention or limited feature sets cannot do.
But delivering impressive specs is only the first step. Greenfield has been hard at work making sure that the silicon will actually do all the things it promised to do. They wisely verified the entire chip set's functionality on an enormous FPGA-based test bed a year before silicon was available. While it was a rather costly exercise, it paid off in the long run by saving several design spins and associated costs. Having a working prototype also parallelized software development, allowing a near-simultaneous bring-up of LVL7 and IP Infusion software on working boards within 2 weeks of silicon arrival.
And speaking of software, Greenfield has made sure that designers would have a good set of development tools to work with. This includes a fully-validated software API suite that's been tested for compatibility with two 3rd party protocol stacks (LVL7 & IP Infusion) in a real network environment.
You have your choice between the Linux and VxWorks operating systems. But if neither of these strike your fancy, the chip set's functional APIs are independent of the underlying hardware, operating system or protocol stack, allowing for easy porting to new OS.
In summary, Greenfield is to be congratulated for delivering a chip set that should raise the bar for price, performance, and flexibility in networking silicon. And to their credit they've done so without much of the hoopla, hype, and fanfare that usually accompanies a roll-out of this magnitude. Rather than tout their wares ahead of time, they have waited for verification of all features working completely before "uncloaking" their company and their chips simultaneously.
Given the detailed information I received under NDA that vouches for their ability to deliver on their promises, the chips earn a respectable 2-saltshaker Vapor Index Rating. The only thing that keeps it from having a half-saltshaker lower rating is the sheer complexity of the parts that could be hiding a bug or structural problem -- despite Greenfield's exceptionally rigorous verification process. Only time will tell if they really have it right but, for the moment, it looks like they have delivered a winning combination of price, performance, and versatility.
The G120, G525, and G750 are all available now. The G525 and G750 are both priced at $495 each in small quantities.
|