What Cisco is up to--From todays EETimes--
eet.com
Cisco promises 10-Gbit upgrade to GSR router family
By Loring Wirbel
EE Times
(12/15/99, 11:38 a.m. EDT)
SAN JOSE, Calif. ? After laboring for several quarters, Cisco Systems Inc. is finally showing the results of its work on packet-over-Sonet interfaces, putting 10-gigabit/second line cards into trials and debuting a new model of its popular Gigabit Switch Router (GSR) family.
The GSR 12016 system, with 10-Gbit/s line cards and a 320-Gbit/s aggregate capacity, is Cisco's answer to the number of core-router startups that have emerged in the past year.
Cisco is also clarifying the directions it will take for terabit/second performance in core backbones. The company is developing a 256 x 256 crossbar switching fabric chassis that will exist as a separate system from its GSR routers. This fabric will replace the internal fabric inside the new GSR 12016 router and other members of the family, and will allow as many as 16 GSR 12016 routers to be linked into a 5-Tbit/s aggregate route platform.
Graeme Fraser, vice president and general manager of Cisco's optical internetworking business unit, said the new terabit system is not a clustered suite of routers, but rather a true scaled system in which route protocol processing and route table lookup can expand to handle a network of 16 GSR routers. Combined with the new 10-Gbit line cards, the new networked switching fabric of routers can support up to 6.4 billion packets per second.
The new line cards will use special very-short-reach optics, implemented as parallel vertical cavity surface emitting lasers (VCSELs), in order to link the GSR 12016 with the Monterey Networks optical cross-connect, or with dense wave division multiplexing (DWDM) equipment from external vendors. It is likely that the network of GSR routers with an external switching fabric will employ these in-building very-short-reach optics as well.
Fraser said that Cisco had examined multi-dimensional router architectures being studied by many startups, including toroids, hypercubes and butterfly topologies, and concluded they could not scale in as straightforward a manner as an external switching fabric. Cisco has not closed the door on advanced topologies, Fraser said, but is using the switching fabric option as the most cost-effective next-generation move to terabit routers.
Cisco's startup competitors consider it a major admission that Cisco has validated multistage routing, the need for terabit performance, and the value of fiber interconnect between chassis within a Point of Presence or central office. Michelle Rae McLean, director of strategic marketing at Pluris Inc. in Cupertino, Calif., said that Cisco had never before admitted that these aspects were important in router design.
"It's very true that it's easier to implement an external fabric, but without multidimensionality, it's a whole different ball game in terms of scaling," McLean said. She said that Pluris' seven-dimensional multistage fabric can scale in theory to 184 Tbits/s, though such systems will not ship for many quarters. Nevertheless, the scaling potential is much greater than for Cisco's GSR 5-Tbit system, she said.
McLean agreed with Fraser's observation that adopting routing protocols and table look-ups to multi-dimensional architectures was a hard problem. Implementing parallelized BGP4 and similar protocols is one of the hardest problems in designing such routers, she said. Independent scaling of the routing table processing must be accomplished, and that task requires dedicated hardware as well as parallelized software, she said.
Cisco's new 12016 is the top-of-the-line in the GSR family, and scales to 320 Gbits/s. The 16-slot chassis has an internal 64 x 64 switching fabric with hardware scheduler, which is replaced with a Terabit Scalability Module when the external 256 x 256 fabric is used. The router uses a dedicated Route Processor card to handle table lookup, which is essentially the same processor used in the earlier 12008 and 12012 GSR routers.
Cisco is using several higher-layer technologies to improve throughput. On each 10-Gbit line card, a feature called Cisco Express Forwarding (formerly termed the Forwarding Information Base) hands off all advanced forwarding decisions on a distributed basis, leaving the RP card free to deal only with control plane issues. In addition, dynamic packet transport utilizing spatial reuse protocol allows packet traffic to be multiplexed on fiber rings, which takes advantage of the concatenation possible in the OC-48c and OC-192c technologies Cisco has implemented.
Cisco went through two generations of semiconductor company acquisitions to handle the two Sonet rates for packet over Sonet. For its OC-48c (2.5-Gbit/s) line cards, Cisco relied on technology from Canadian startup Skystone Corp. And its acquisition of Stratum One Communications Inc. earlier this year was central to completing its OC-192c line cards. Fraser said the ability to support concatenated versions of Sonet rates was critical, since it allows more multiplexing of slower services on broadband Sonet pipes. Qwest Communications International Inc. is conducting nationwide tests of the 10-Gbit cards, though Fraser said they are still many months away from volume production.
Cisco's strategy of interconnecting routers through an external switching fabric, as well as its strategy of using short-haul optics to link routers, DWDM multiplexers, and optical cross-connects (OXCs), indicates that Cisco executives do not think Points of Presence and central offices of the future will be dominated by one large route/switch device. Distributed hardware platforms will be central to all Cisco platforms moving forward, Fraser said. The Monterey OXC, re-branded as the Cisco ONS 15900 Wavelength Router following Cisco's acquisition of Monterey, will be a key element of a network linking routers, IP/ATM switches, and data-aware add-drop muxes from the Cerent portfolio, into a long-haul DWDM backbone.
The launch of "baby gigarouters" such as the M20 from Juniper Networks Inc. earlier this month has raised an important issue as to whether core routers and edge routers are melding into one very high-performance architecture. Fraser said the notion of melding makes sense in terms of absolute speed, as a 40-Gbit/s router may soon be required at the edge of the public network. But there will always be a difference in the mix of aggregation and forwarding functions required of core routers and edge routers, he said.
The edge also is an area where time-division-multiplexed Sonet nodes will continue to play an important role, Fraser said. Even as Sonet disappears as an independent layer in 10-Gbit cores, Cisco's interest in the Cerent TDM Sonet platforms will continue to be critical, Fraser said. |
