Scalability issues in metro/access networks
SPECIAL REPORTS: Optical Data Networks
Should you get into the ring with resilient packet ring technology, Ethernet switching, or SONET?
KANAIYA VASANI and MANNIX O'CONNOR, Lantern Communications
The downturn in the economy has brought with it a major paradigm shift in business strategies. The previous land-grab mode of operation-get market share at any cost-has given way to the survivor mode-make the most of limited resources. In some sense, it is back to basics, a focus on fundamental business values such as return on investment.
The "throw bandwidth at the problem" approach to removing the metro network bottleneck is not getting many takers in this economic environment. Communications service providers must deploy technologies that maximize the number of profitable services offered over their network infrastructure. They also need to bring services to market as quickly as possible.
These business drivers dictate that metro network infrastructure must scale with growth from initial deployment to tens of thousands of customers. Successful service offerings may require a provider to turn up hundreds of customers each day.
Scalability is an issue both from a service perspective and from a network perspective. Ethernet services are gaining broad popularity because of their inherent scaling advantages. How does the emerging resilient packet ring (RPR) technology compare to incumbent architectures based on Ethernet switching and SONET add/drop multiplexers (ADMs) in terms of network scalability and efficiency?
Simply defined, an Ethernet service is any data service offered via an Ethernet interface-a 10-Mbit/sec, 100-Mbit/sec, or 1-Gbit/sec Ethernet port. A key difference between Ethernet services and legacy data services such as leased lines, frame relay, or ATM is the scalability of the service interface.
With legacy data services, physical interface requirements vary with the speed of the service. Thus, hardware required for a T1 (1.54-Mbit/sec) service is completely different from that required for DS-3 (44.736-Mbit/sec) or OC-3 (155.52-Mbit/sec) services. With Ether net service, on the other hand, a service provider can drop a Fast Ethernet (100-Mbit/sec capacity) or Gigabit Ethernet (1,000-Mbit/sec capacity) port to a subscriber once and upgrade many times, without additional truck rolls beyond the initial installation. Bandwidth and other service changes can be administered remotely, simplifying and accelerating service provisioning.
Ethernet services are widely viewed as an offering that holds promise for rapid acceptance in the marketplace. The question remains what infrastructure can cost-effectively scale to meet this demand.
RPR is an emerging network architecture and technology designed to meet the requirements of a packet-based MAN. Unlike incumbent architectures based on Ethernet switches or SONET ADMs, RPR approaches the metro bottleneck problem with a clean slate.
In the past few years there have been fiber ring deployments in most metro areas. The challenge for service providers is to tap into the latent capacity available on these fiber rings and carve out as many profitable, revenue-generating services as possible. This problem of effectively managing a shared resource is typically solved at the media access control (MAC) layer of the protocol stack.
RPR is a new MAC protocol designed for metro fiber ring networks. It is being standardized by the IEEE within its 802.17 working group. The standardization effort has received widespread industry support with participation from more than 90 companies from different parts of the world. The first draft of the standard is expected in the first quarter of next year.
By creating a MAC protocol for ring networks, RPR attempts to find a fundamental solution to the metro bottleneck problem. Other solutions attempt to make incremental changes to existing products but do not address the basic problem-SONET ADMs and Ethernet switches do not address the need for a MAC layer designed for the metropolitan-area environment. SONET employs Layer 1 techniques (point-to-point connections) to manage capacity on a ring. Ethernet switches rely on Ethernet bridging or IP routing for bandwidth management. Consequently, the network is either underutilized in the case of SONET or non-deterministic in the case of Ethernet switches. RPR has several unique characteristics that make it an ideal platform, in terms of scalability and efficiency, for delivery of data services in metro networks.
In an RPR network, the entire ring bandwidth is available for use at all times. Unlike SONET, statically configured TDM channels are not defined for data service within the ring. With RPR technology, each node on the ring has visibility into and access to the entire capacity of the ring. Therefore, the entire ring bandwidth is used efficiently for any customers requesting service based on their contracted service level agreements (SLAs).
In a metro network based on traditional Ethernet/packet switches, each node has knowledge and use of capacity on its immediate links only, not the entire network. Consequently, the best a node can do is to optimally utilize capacity on its immediate links. But that presents a problem, because traffic that was admitted at one node may run into congestion at the next node; network nodes have no visibility into network utilization at any other node.
In the world of best-effort services that is not an issue, because traffic may be discarded in a congested environment. But if carriers wish to deploy revenue-generating services based on SLAs, it becomes a critical issue. Ethernet switches overcome this problem through network over-provisioning. On average, only 40% of the capacity on a network of Ethernet switches can be used to deliver guaranteed services.
On RPR networks, on the other hand, close to 100% of the capacity on the network can be used to deliver guaranteed services with SLAs.
There is a second aspect to RPR that provides better scalability. RPR protocols incorporate destination stripping to reclaim bandwidth. This feature requires that the destination node remove the packet from the ring. Imagine the numbers on the face of a clock superimposed on 12 nodes of a metro ring. If a packet enters a ring at node 1 and is destined for node 4, then on the balance of the ring, nodes 5 through 12, no bandwidth is consumed by this traffic. That implies that node 6 could be sending traffic to node 9, while at the same time node 1 is communicating with node 4. Depending on the nature of the customer source and destination pairs, the theoretical maximum bandwidth equals the number of nodes times the data rate of the ring. In this way RPR rings scale many times beyond their basic data rate.
It is interesting to compare the packet ADM architecture of RPR devices with the packet switch architecture of Ethernet switches. As shown in Fig ure 1, a metro network built with Ethernet switches consists of nodes connected by point-to-point links. Network traffic gets queued and scheduled at every intermediate node between the source and the destination. That poses a serious scalability issue. Each node now has to process traffic coming in from the network at line rate. Packet processing technology at each node may be able to handle the lower rates of 1 Gbit/sec or 2.5 Gbits/sec. But when the network scales to 10 Gbits/sec and beyond, this approach breaks down.
RPR devices on the other hand implement the notion of a transit path. At each node, traffic that is not destined for the node simply passes through; it does not get queued and scheduled. As shown in Figure 2, the MAC entity on each node performs three functions: "add" or insertion of subscriber traffic from the node, "drop" or removal of traffic destined for a subscriber on the node, and "pass" or direct transfer of transit traffic from one network link to another. The transit path effectively becomes a part of the transmission medium and makes the resilient packet ring behave as one continuous medium shared by all the RPR nodes. Because a packet ADM node does not process transit traffic, the packet ADM architecture can scale more easily to higher data rates.
Another attribute unique to RPR is active and dynamic management of bandwidth on the ring. In a network with dynamically changing traffic patterns, which is typical in any packet network, the only way to optimize network utilization without discarding traffic is to have an active feedback mechanism built into the network. The feedback mechanism informs the traffic sources of the capacity available on the network so that the sources can adjust the rate at which they inject traffic into the network.
The RPR MAC is bandwidth-aware. The MAC entity on each node monitors the utilization on its immediate links and makes that information available to all the nodes on the ring. Each node can then either send in more data or throttle back. This efficient use of bandwidth enables the ring to scale beyond 95% of its total capacity.
Ethernet switches or SONET ADMs do not have any bandwidth-management capabilities and hence cannot maximize network utilization.
A common complaint from data service customers is that it takes too long for carriers to provision services. Activation times on the order of six weeks to six months for DS-1 and DS-3 services are quite common, with services at OC-3 rates and higher taking even longer.
A significant portion of this delay in service activation can be attributed to the underlying SONET infrastructure and its circuit-based provisioning model. Creating an end-to-end circuit takes many steps. First, the network operator identifies the circuit's physical endpoints to the management system. The operator must then configure each node within the ring for all the required pass-through and add/drop connections. This provisioning operation is time- and labor-intensive.
Newer SONET systems automate some of the provisioning steps. But the network operator still needs to perform traffic engineering manually to optimize bandwidth utilization on the ring. The operator must be aware of the network topology, the traffic distribution on the ring, and the available bandwidth on every link traversed by the circuit.
Service provisioning on a network of Ethernet switches is slightly better because they do not require provisioning of circuits through each node. Provisioning, however, still happens a node at a time. In addition, if providers wish to deliver SLAs over the network, the operator still needs to manually traffic-engineer the network.
An RPR system, by contrast, offers a very simplified service model. In an RPR system, the ring functions as a shared medium. All the nodes on the ring share bandwidth on the packet ring. Each node has visibility into the capacity available on the ring. Pro visioning a new service is therefore far simpler. There is no need for a node-by-node and link-by-link capacity planning, engineering, and provisioning exercise. The network operator simply identifies a traffic flow and specifies the quality of service it should get as it traverses the ring.
After all the comparisons to Ethernet switches and SONET ADMs, it may appear that RPR intends to compete with two well-entrenched, time-tested technologies. That's not the case, however. Rather, RPR is complimentary to both SONET and Ethernet. Both SONET and Ethernet are excellent Layer 1 technologies. SONET was designed as a Layer 1 technology, and Ethernet has evolved into one. Through its various incarnations, Ethernet has transformed from the carrier sense multiple access/collision detection shared media network architecture, to a full-duplex, point-to-point switched network architecture. Most of the innovation in Ethernet has been focused on its physical layer, increasing the speed at which it operates. The MAC layer has been largely left untouched and is practically irrelevant. The portion of the MAC layer that continues to thrive is the MAC frame format.
RPR is a MAC protocol and operates at Layer 2 of the open systems interconnection (OSI) protocol stack. By design, RPR is Layer 1 agnostic, which means that RPR can run over either SONET or Ethernet. RPR en ables service providers to build more scalable and efficient metro networks using SONET or Ethernet as physical layers.
These comparisons are between RPR and proprietary extensions to SONET and Ethernet used in vendors' products, be it SONET ADMs or Ethernet switches. RPR in conjunction with Ethernet or SONET, on the other hand, offers a standards-based approach to building highly scalable metro networks.
Kanaiya Vasani is director of marketing and Mannix O'Connor is director of business development at Lantern Communications (San Jose, CA).