Bandwidth is Dead. Long Live Data Logistics

In a recent post on the Internet Society’s blog entitled “Bandwidth is Dead. Long Live Latency,” Jason Livingood, vice president of Technology Policy, Product & Standards at Comcast, argues that we are in a “post-gigabit era.” His position is that the evolution of network metrics should focus on minimizing end-to-end transmission latency.

Livingood argues for a focus on Quality of Outcome (QoO): “Unlike traditional methods that emphasize speed or bandwidth, QoO focuses on what really matters to users—their actual experience with applications like video conferencing, gaming, and streaming.”

Livingood’s proposed approach to QoO focuses entirely on telecommunications (synchronous end-to-end data transmission), spanning both wide and local area networks. Rather than focusing on “artificial” metrics such as theoretical network bandwidth and uncongested end-to-end latency, he wrote, “An alternative approach is emerging, shifting away from artificial measurements that are indirect proxies of Internet quality to directly representative or predictive measurements.

“These approaches can leverage an artificial intelligence agent running in customer premises equipment (CPE) in a home network to observe:

• the performance of the access network and the Internet beyond
• the user’s local wired and Wi-Fi network, including down to a per-device and per-application level.”

Livingood offers four examples of “end user experience”:

• FaceTime quality
• YouTube quality
• Buffering in Smart TV streaming
• Bandwidth-limited software downloads

and suggests dual-queue networking as a solution:

“Internet service providers can now take steps to optimize network responsiveness. This can involve deploying newer active queue management algorithms in the network and CPE and implementing IETF L4S and NQB, known as dual-queue networking.”

This analysis ignores techniques that combine storage and computation with data transmission to improve the quality of many network-constrained applications. Many such latency hiding techniques avoid the technically challenging problem of minimizing end-to-end transmission delay. Latency hiding is the principle that originally motivated the emerging field of data logistics.

Data logistics emerged in the deployment of FTP “mirror sites” in the 1980s. These use content replication to optimize the distribution of data being published over the network. In the 1990s, this gave rise to the Web caching movement, which eventually evolved into the modern content delivery networking industry, a predecessor to the cloud.

The store-and-forward architecture of the Internet Protocol (IP) was itself an early step in the use of asynchrony to optimize end-to-end data transmission. The benefits of asynchrony in the global datagram delivery service are then reconciled with the needs of synchronous telecommunications applications through the use of timers in the Transmission Control Protocol (TCP) at the transport layer. The result is a service which delivers imperfect synchronous connectivity.

The buffering available at IP intermediate nodes (routers) relaxes the stringent timing requirements that were previously imposed by digital telephony. From an early point in the evolution of the Internet, there was an understanding that supporting mass media using store-and-forward also requires some form of efficient point-to-multipoint communication. IP multicast was an attempt at one such mechanism. But all early such efforts to support data logistics were met with pushback from the traditional end-to-end telecommunications-based networking community or by a lack of widespread adoption.

In the 1990s Active Networking was a concerted research effort funded by the Defense Advanced Research Projects Agency. Active Networking sought to increase the application of storage and processing to the generalization of networking beyond simple end-to-end data transmission. A well-funded attempt at using latency hiding in a public network was information centric networking. Another approach to convergence of storage and networking at the turn of the century was logistical networking, which sought to combine storage, networking, and computation in a converged model of wide area infrastructure.

The understanding that minimization of end-to-end latency requires latency hiding techniques is widely accepted. The 2004 paper “Latency lags Bandwidth,” by Turing Award recipient David Patterson, discussed the use of data replication (storage) and synthesis (processing). The seminal 1982 paper “End-to-End Arguments in System Design,” by Jerry Saltzer, David Reed, and David Clark has been used for decades to argue that introducing any functionality beyond packet forwarding into the network is counterproductive and a threat to the scalability of network deployment and operation. This seems to set up a paradox: latency hiding within the network is both required and forbidden.

A resolution of the paradox is easy to find: the end-to-end arguments are not a logically valid argument, but more a rule of thumb. Exceptions have been made by the authors, for reasons ranging from security and performance in the original paper, to a later “carve out” for active networking. Nonetheless, the proponents of the end-to-end arguments credit it as a key reason for the continued stability and explosive growth of Internet connectivity. They turn a blind eye to cases in which it does not apply, in an epic case of confirmation bias.

A more recent formal analysis of the “deployment scalability” of layered systems provides an alternative to the end-to-end arguments that explains its power in many cases where it does apply, but which differs in other cases. The principle of “Minimal Sufficiency” explains that community standards which are logically weak will tend to have more deployment scalability, but it allows for the minimal strengthening of standards when necessary to achieve specific goals. Adding storage and processing to the functionality of the network is an example of “logical strengthening.” Minimal Sufficiency explains that logical strengthening should take the weakest form that still meets necessary goals. This stands in contrast to the end-to-end arguments, which are widely interpreted as forbidding any increase in network functionality. This has become a point of orthodoxy among proponents of end-to-end.

In this new analysis, an important problem with the previous efforts to increase the functionality of the public network is that they tended to be too logically strong. Researchers who chose not to be constrained by the end-to-end arguments did not restrain themselves as suggested by Minimal Sufficiency. This “all or nothing” approach does not take account of the fundamental tradeoff between logical strength and deployment scalability expressed in the hourglass theorem. We are overdue to consider Patterson’s replication and synthesis strategies in a manner that is guided by Minimal Sufficiency. It is time for the era of telecommunications (end-to-end transmission) to give way to the area of data logistics (converged storage, transmission, and computation).

In the absence of the shared public infrastructure that supports data logistics, alternative approaches have been pursued. Private content delivery network infrastructure, in the form of backbone networks and distributed datacenters, are not constrained by cost and engineering difficulty. They serve only paying customers in the most developed components of the global Internet. At the same time, telecommunication operators are trying to segment the public Internet into “lanes,” segregating traffic (such as L4S) requiring the lowest possible latency (for example, teleconferencing) from traffic that is less sensitive to delay (such as file transfer).

There is a long history of failed attempts to introduce quality of service (QoS) guarantees within the public Internet. These efforts have been focused on guaranteeing a minimal bandwidth throughout the lifetime of a transport layer connection. An important problem with such approaches is that the reservation of bandwidth within an otherwise best-effort total is a strong guarantee which reduces the deployment scalability of the network. Proponents of creating a “low latency lane” such as L4S seem to think this is not a form of QoS guarantee.

Fundamentally, Quality of Outcome has as little to do with end-to-end transmission latency as with end-to-end bandwidth, because end users (and endpoints) don’t care where the response comes from. What they do care about is low interaction response times, of which transmission latency is just one component, and the availability of rich services that meet their needs at a low cost with integrity. It doesn’t matter to them if they are communicating with cloud servers over an ultra-low latency network connection or if they are communicating with a proxy in their end network or integrated with the core network. In other words, the latter approach utilizes latency hiding using the techniques of data logistics. However, to an executive in one of the largest Internet service providers, connecting end users to the cloud at low latency is the natural, and perhaps the only, solution.

See also motivated reasoning, path dependence, double mindedness, technology silos.

Micah D. Beck (mbeck@utk.edu) is an associate professor at the Department of Electrical Engineering and Computer Science, University of Tennessee, Knoxville, TN, USA.