networkZONE Products for the week of January 6, 2003
Marvell Says…
A Logical Progression - Marvell Adds Larger Redundant
Crossbar Switch Fabrics To Its Prestera Architecture
The Prestera Device, Integrating 72 SERDES on a Single Chip,
Enables OEMs to Enter and Expand Into the Enterprise and Aggregation Switch
Markets
Marvell has extended its Gigabit Ethernet switching leadership by offering
the industry's highest performance, most integrated crossbar switch fabric.
The Prestera device integrates seventy-two 3.125 Gigabit per second (Gbps)
serializer/deserializers (SERDES) on a single chip, enabling OEMs to enter
and expand into the enterprise and aggregation markets with a robust 360
Gbps switching system. The Marvell Prestera-FX crossbar switch fabric leverages
the same proven SERDES expertise that is the driving technology behind the
Company's industry-leading Alaska X 10 Gigabit physical layer transceivers.
OEMs can now take advantage of this increased SERDES integration for even greater system scalability and performance. For example, the Prestera-FX crossbar switch fabric provides OEMs with the ability to develop 12-blade chassis systems that scale up to 576 Gigabit Ethernet (GbE) ports with 48 GbE port line cards for wiring closets, or 48 10-GbE ports for backbone connectivity. Used in conjunction with Prestera fabric adapters and packet processors, the Prestera-FX crossbar switch fabric enables advanced chassis features such as redundancy, hot-swap and centralized management capabilities.
"Marvell is clearly extending its leadership position in providing high-density, high-performance switch fabrics," said Gary Smerdon, Vice President of Marketing for Marvell's Switching Products Group. "Complete chassis solutions based on the Prestera packet processors, switch fabrics and Alaska PHY transceivers enable OEMs to quickly develop industry-leading platforms for enterprise and service provider networks."
The Prestera-FX Crossbar Switch Fabric
Marvell offers two crossbar switch fabric solutions, providing manufacturers
with scalable building blocks for high-performance chassis systems. Both
Prestera devices take advantage of a cell-based, output queue architecture
that supports up to 144 Gbps non-blocking performance. Variable cell size,
another Prestera-FX feature, guarantees non-blocking throughput at all packet
lengths. The Marvell solution also resolves fabric congestions by using
in-band flow control messaging. In addition, hot swap and redundancy capabilities
add reliability for advanced chassis systems based on the Prestera-FX switch
fabric.
The Prestera-FX9210 Crossbar Switch Fabric
The Prestera-FX9210 crossbar switch fabric is a 12-port switch fabric that
integrates 72 SERDES ports running at 3.125 Gbps, providing a net bandwidth
of 360 Gbps through the fabric. With the ability to support a network load
in excess of 700 million packets per second when fully utilized, the device
can support 144 Gigabit Ethernet ports running at full-wire-speed or as
many as 576 Gigabit Ethernet ports in a typical system. Despite the large
number of high-performance SERDES, the Prestera-FX device only consumes
18 Watts of power, thus taking the lead in low-power consumption.
The Prestera-FX9110 Crossbar Switch Fabric
The Prestera-FX9110 crossbar switch fabric is a 9-port switch fabric that
integrates 54 SERDES ports running at 3.125 Gbps, providing a net bandwidth
of 270 Gbps through the fabric. This product supports a maximum of 108 Gigabit
Ethernet ports running at full-wire-speed, and consumes approximately 13
Watts of power.
analogZONE Says . . .
Even if Marvell's Prestera architecture had not won one of our Product of the Year awards for their Prestera switch architecture, I'd be writing up the addition of two new crossbar switches to their chip family because it fills a void I've been staring at for the past several months. If you're unfamiliar with the Prestera EX/MX chip sets, feel free to take a look at the full review I did on them last year.
Prestera is a versatile switching solution that Marvell has applied with equal success to chip sets for the workgroup, enterprise, and metro markets, but until now there has only been a limited amount of expansion possible using the FX series of switch fabric adapters. As I explained in my previous review, the entire Prestera switch family uses a deterministic architecture where packets are broken into fixed-length cells for storage and manipulation within a shared-memory buffer on the switch controller element. Incoming packets are queued by the blade's local packet processor and buffered on its input queue. If traffic is destined for another controller, is shipped between switch elements via multiple lanes of 3.125 Gbit/s binary SerDes channels (actually 2.5 Gbit/s worth of data, plus overhead for the 8B/10B channel coding.)
The previous generation of controllers did a nice job of connecting ports across a passive backplane, but only allowed for limited levels of expansion, and the high-capacity serial channels seemed to beg for some sort of high-density switch fabric that would allow you to build much larger, more complex connect schemes, and chassis-based products. Unfortunately, the earlier FX controllers did not support chassis-based designs with more than a half-dozen blades.
Happily, Marvell's two new switch fabric chips (the 72-channel FX9210 and 54-channel FX9110 ) solve this problem and allows you to hook up larger numbers of fabric controllers using the same simple passive backplane technology. Perhaps, more important, the new crossbar elements bring several important carrier-class reliability features to the product line.
With 72 SERDES channels, the Prestera-FX9210 crossbar switch fabric provides 12 interconnect ports between switch port controllers with a net bandwidth of 180 Gbit/s Full-duplex.) Its 9-port cousin, the FX9110 crossbar has 54 SERDES channels with a full-duplex bandwidth of 169 Gbit/s. Actually the 9210 sports 13 ports, with the extra one being used as a control, service, and monitoring channel. Likewise, the 9th channel on the 9110 serves the same purpose.
The new switch fabrics are big improvements over their previous versions, with several important new features that will help make the design of equipment with carrier-class reliability much easier. For one thing, the chips have built-in error correction within the switch fabric. A CRC-type code is inserted as part of the overhead and is transparent to actual traffic, and allows detection and correction of errors without any intervention from the control processor.
Equally important is the newly-developed hot-swap capability and support for redundant control CPUs that can be changed at will. But perhaps the most dramatic feature is the support for one-for-one redundancy, allowing the construction of fault-tolerant switch fabrics that are comparable to PMC-Sierra's carrier-grade products, but more highly integrated.
This redundancy also permits you to use the fabric to perform load balancing across your switch. Marvell was a bit vague on the details here, but from what I understand you can perform load balancing across both sides of the switch by coordinating the egress queues on both sides of the redundant fabric to offload one side when it is getting full, and use redundant switch fabric to carry extra capacity. This coordination is done via in-band signaling with control cells. From what I was able to pry from Marvell, this allows traffic to be load balanced with no out of order packets.
As-always, the chipset has on-chip replication for multi-cast and snooping/sniffing. I've been pleased at how they have used this common architecture to help create a powerful set of OEM and 3rd-party development tools and reference designs, and expect that whatever new features are now available will show up in the latest APIs, as well as reference designs, supplied by RadLAN and other 3rd-party development outfits.
While I was a bit put off by the lack of detail
on the hot swapping and signaling used in the two switch fabrics, I'm familiar
enough with the product line to take Marvell's claims as the truth - at
least for the moment.
| |