The FAST Copper Project

Motivation

KEY IDEAS AND TECHNICAL CHALLENGES

Traditionally, DSL broadband access networks have been analyzed by viewing each twisted pair as a separate communication channel, independent of other twisted pairs in the same binder cable, with a fixed pipesize supporting circuit-switched voice traffic. The key to realizing the vision of ubiquitous, readily deployable, and truly broadband access networks is to dynamically optimize the resources in the dimensions of Frequency, Amplitude, Space, and Time, in the multiple-inputmultiple- output communication environment of DSL across multiple layers in the protocol stack. Therefore, there are two shifts of mentality that underlines the wide range of activities in FAST Copper:

The first key idea is that, instead of holding the traditional view that each twisted pair is an independent channel, we model a bundled cable of twisted pairs as one aggregate multi-user communication system. Multiple users compete against and cooperate with each other in this system. We can explicitly take into account the crosstalk effects (both near-end and far-end) that currently form the data rate bottleneck, and to exploit potential cooperation in sharing limited resources

The second key idea is that today’s traffic over broadband access, including voice, data, and video, are predominately supported by packet switched IP. We can exploit the burstiness of the application traffic through aggressive statistical multiplexing, with admission control, traffic shaping, scheduling, and priority queuing mechanisms to ensure the desired tradeoff between the number of application flows supported and the Quality of Service (QoS) attainable. There are two major bottlenecks to DSL broadband access today: attenuation and crosstalk. We will see solutions from the “Space” dimension of the project to tackle the problem of attenuation, and solutions from the “Frequency”, “Amplitude”, and “Time” dimensions to tackle the problem of crosstalk. Note that we have not even brought in factors such as wider bandwidth and multiple twisted-pairs.

Frequency. In the physical layer, new techniques can be developed based on improving spectral utilization, mitigating multi-user interference, and exploiting multiuser cooperation. Through dynamic adaptation and utilization of frequency spectrum, such as power control, bit loading, or vectored transmission, Dynamic Spectrum Management (DSM) allows maximum flexibility in allocating rates among competing flows, achieves much higher total data rates, and extends the reach of broadband access.

Time. FAST Copper also leverages the potential for time division multiplexing based on the application layer burstiness of data traffic from and to the end hosts. In most communication-theoretic investigations, it is assumed that there is always an infinite backlog of bits that need to be transmitted per user, thus taking out the latency considerations and the temporal dimension. By jointly considering the application layers, burstiness of the required bandwidth provides another degree of flexibility of statistical multiplexing along the temporal axis.

Space. When building robust and efficient broadband access networks, two issues are particularly important: how can a hybrid fiber/twisted pair architecture be designed to utilize the best of fiber-based and copperbased communication potentials, and how can a logical topology be designed to offer fast-recovery after natural failures or malicious attacks?

Amplitude. We propose to install active ‘amplitude control’ mechanisms to shape the flow intensities at the edge to provide different QoS classes through admission control and dynamic bandwidth allocation. At the same time, a network management system constantly probes, measures and monitors the cable and its environments, receives data rate requests from user terminals, and periodically shapes the rate each user is allowed to transmit and receive per time frame. In summary, by modeling the whole binder of copper wires as one multi-carrier interference channel, with resources ranging from the physical layer to the application layer, we can dynamically optimize over Frequency, Amplitude, Time, and Space, in a stable, robust, and complementary way. Collectively, these four degrees of freedom offer many exciting opportunities to make tangible practical impacts. At the same time, progress in the project come from solving important problems in the fundamental research disciplines of information theory (multi-carrier interference channel), signal processing (multi-user transceiver design), optimization theory (nonconvex and coupled problems), graph theory (survivable tree topology design), stochastic theory (processor sharing and queuing networks), distributed control (feedback control at different timescales), and network protocol design (resource allocation and functionality allocation).

ARCHITECTURAL ISSUES IN BROADBAND ACCESS

A. Functionality Allocation

Increasing data rate by 10 times (or more) over twistedpair already presents tremendous technical challenges. We need to significantly improve both the digital signal processing algorithms in the physical layer and the architecture/protocol design methodologies in the “upper” layers. Even more challenging is the need to carefully investigate the coupling effects across multiple modules and across network elements, so that end-user experience over broadband access networks is enhanced. Indeed, one of the most important aspect of FAST Copper, or any broadband access network design, is on the access network architecture. Following the notion of “architecture first”, we open the technical discussion on architectural choices. Architecture here refers to functionality allocation: which functional module and network element does what, and how to connect them. Functionality allocation is often more influential, harder to change, and less quantitatively understood than any specific resource allocation scheme. Metrics of measuring the pros and cons of designs of functionality allocation often remain fuzzy today, and are drawn from a combination of performance metrics, cost and complexity metrics, and Network X-ities (e.g., evolvability, scalability, manageability, diagnosability, optimizability) metrics. Recent results in “Layering as Optimization Decomposition” have offered a useful framework for layered network architecture. There are unique challenges in architectural issues in broadband access networks, as elaborated below.

B. Horizontal and Vertical Decompositions

There are two types of functionality allocations:

First, we use the term “horizontal decomposition” to refer to the geographic distribution of control into various network elements, e.g., in Figure 1, from households to remote terminals and central offices (represented by circles), to larger central offices (represented by ovals), and to backbone acquisition, distribution, and video servers (represented by cylinders).

Second, we use the term “vertical decomposition” to refer to the modularized design into a protocol stack. Ideally for performance optimization, a fully integrated and joint design would be best. However, for many reasons such as evolvability and manageability, modularized design is necessary. Horizontally and vertically decomposable modules has a variety of coupling relationships, depending on their timescales and target service models. As an example of such couplings, consider the functionality of error control and recovery. We can choose from the following: hop-by-hop forward error correction or feedback based Automatic Repeat reQuest, or end-to-end connectionless resilient-UDP or connectionoriented TCP, or coding over packets at the application level. Which combination to choose from involves both vertical decomposition (e.g., physical layer or application layer) and horizontal decomposition (e.g., hop-by-hop or end-to-end). In the horizontal decomposition, the problem of where to place video servers involves the tradeoff between response time and scalability, closer to the customers, faster the response but lower the scalability. Similarly, the problem of where to place distribution servers defines the boundaries of multicast groups. An even larger issue is on “how big should the access networks be”. The answer depends on the tradeoff among a variety of factors, from reliability of access tree and feasibility of big switches to complexity of backbone network and ease of network management.

In the vertical decomposition, the four dimensions of F, A, S, and T are all coupled, sometimes in unexpected ways. As a thought experiment in the extreme, we can think of each of these dimensions as capable of tackling crosstalk: through spatial division multiplexing, time division multiplexing, frequency division multiplexing, and rejection of all flows that would transmit at the same time. Obviously, these are extreme measures, and tradeoffs are necessary in any design. But they do highlight the fact that these four dimensions are not uncorrelated degrees of freedom.

Topology design (e.g., placement of various servers and schedulers) determines feasibility of control in the time and amplitude dimensions, since excessive propagation delay due to geographical distance may lead to instability of scheduling at the packet level and impossibility of admission control of short flows.

Topology design also determines the crosstalk channel gains, which are the parameters to spectrum management algorithms in Frequency dimension and limit the best rate regions attainable.

Time and frequency are clearly coupled design freedoms. Each spectrum management or scheduling can mitigate crosstalk. As will be shown, advanced versions of dynamic spectrum management provides a convex rate region, thus rendering the simple time division scheduling inferior. The timescale of spectrum management depends on the speed of convergence (time complexity) of spectrum management algorithms, and that of scheduling depends on the granularity of schedules: anywhere from packet level schedules to flow level ones. Furthermore, the capability in the time dimension depends on the properties of spectrum management in the frequency dimension.

Admission control in the amplitude dimension clearly depends on the rate regions that are assumed to be attainable by the underlying spectrum management and scheduling algorithms.

It is important to realize that the timescales of different functionalities can vary significantly. A key point is that more explicit time-scale separations often enables easier decompositions by lowering the “price of modularization”. For example, as shown in Figure 2, problems such as topology design is dealt with on a monthly or yearly basis, whereas time and amplitude-related problems is tackled at the much faster timescale of packet-level or flow-level dynamics. Thus, it does not seem to be of great significance to explicitly consider the space domain in the time and amplitude domain. This is due to the fact that time-scale separation between two dimensions leads to (i) quasi-stationary regime (i.e., nearly constant) and (ii) fluid regime (i.e., nearly average), from the perspective of the dimensions operating with faster time-scale and slower time-scale, respectively.