The exponential growth of Internet Protocol (IP) traffic is well documented, as is the rise of IP technology as the new basis for core network routers. In addition, technologies such as Dense Wavelength Division Multiplexing (DWDM) and Time Division Multiplex (TDM) are decreasing the cost per bit-mile of optical transmission, making the construction of terabit links affordable. The result of these "successes" is a new bottleneck at the core of the communications network.
As new services such as real-time multimedia, including audio and videoconferencing, are accepted into the market and as quality-of-service agreements and differentiated services become available, the problems caused by this bottleneck will continue to escalate. The bottleneck must be circumvented so that carriers can deliver new revenue-based services profitably and efficiently.
Until now, carriers have tended to distribute rack core routers in the network. However, a scaleable central core IP/MPLS (Multi-Protocol Label Switching) router may offer a more efficient and attractive solution to the bottleneck problems than distributed routers. It is becoming clear that there is one key technology needed to make such a centralized router approach viable: a large, fast optical switch fabric.
Distributed vs. centralized
Carriers are demanding scalable routers to handle increasing capacity without having to "forklift upgrade" every few years. The ideal goal would be routers able to scale to capacities supporting a seven- to ten-year product life cycle.
Many of today's core routers have attempted to solve the scalability issue by developing distributed switch fabrics. Distributed switch fabric designs are akin to designing a network within the switch, where packets must traverse multiple nodes to reach the decimation line card. Packets that are received on one line card need to be routed through the switch fabric and could possibly traverse multiple nodes within the switch. As a result, distributed switch fabrics do not effectively solve the networking issues they are designed to address.
The benefit of the distributed architecture lies in its ability to scale. However, as the system grows in size, performance issues related to internal congestion, latency, and loss can also arise.
This point is very significant and is one of the major reasons that all supercomputer vendors have discontinued their N-dimensional mesh-connected parallel processors. The surviving supercomputers architectures use centralized switching fabrics (for example, the IBM Colony Switch for the Department of Energy ASCI Program).
The supercomputer industry went through the distributed-versus-centralized dilemma in the '90s in attempting to build large multiple-processor systems. Various distributed switch architectures were attempted to interconnect the central processing unit (CPUs), but the attempts were found to be overly complex and were eventually abandoned in favor of centralized switch architectures. A similar fate is likely for routers, which will almost certainly need to return to centralized schemes — albeit larger-scale and more sophisticated than the early centralized router architectures — to successfully handle future demands.
The advantages of centralized switch/centralized arbitration architectures lie in their simplicity and include the following:
the performance (latency and throughput) of the switch is predictable and well understood, and it is not dependent on the traffic pattern or number of line cards used. The latency of a packet is relatively constant and is not dependent on the source and destination ports.
scheduling and arbitration is globally optimum, which substantially increases the overall throughput of the switch.
the size and number of interconnections are reduced. In a centralized system, the node is NOT changed when scaling up. Complexity is isolated to the centralized switching fabric.
Both large and fast
Although electronic switch fabrics have scaled remarkably with the capacity demanded of routers so far, today's state-of-the-art electronics require that large electronic-switch fabric designs dissipate too much power and take too much space to be practical.
To date, however, optical switching has either been large and slow, or else small and fast. For example, micro-electro-mechanical systems MEMS (e.g., moving mirror type) devices are now capable of switching up to a few hundred optical signals, but the speed of the switch is limited by the physical mass and momentum of moving parts — restricting the practical speed to milliseconds for single-mirror-per-fiber architectures.
At the other extreme, very high-speed switching devices can be fabricated from Mach-Zender modulators and small waveguide devices, capable of switching light between two fibers in less than one nanosecond. Unfortunately, previous attempts to scale these types of devices to large port counts have proven very difficult due to loss and poor cross-talk accumulation.
The difficult problem is building a fast, large-port-count, readily scalable optical-switching fabric. Chiaro has developed such a large, fast switch based on Optical Phased Array (OPA) technology to support centralized core IP routing applications.
OPA switch technology has no physical limitation to switch in less than one nanosecond, optically, in modules of up to 64 x 64 ports. Practically, when the drive electronics are included, the switching speed has been demonstrated at 25 nanoseconds. As a result, the device suits IP/MPLS core routing applications very well.
Highly available router
Although distributed core-router models can scale effectively, they have difficulty simultaneously achieving the high availability demanded by carriers. Capacity in distributed systems is scaled by interconnecting smaller routers, meaning that the number of managed elements in the network increases. With greater numbers of elements to manage, the risk of failure increases. Achieving carrier-class "five-nines" availability — i.e., 99.999 percent uptime — becomes a costly and complicated affair with distributed routers.
In contrast, centralized routers can more easily attain high five-nines availability.
A large centralized router that combines optical switching and electrical routing can help eliminate the bottleneck that occurs because of the increasing convergence of IP/MPLS, data, and voice traffic in the network. Routers with the OPA switch fabric and based on a centralized architecture have shown in production deployments they can achieve the combination of speed, capacity, scalability, and high availability to meet telecommunications carriers' stringent needs for core routers.
Shekel is founder and general manager of Chiaro Networks, Richardson, Texas.
RSS
RSS| 
Digital| 
Mobile
