AT&T Previews Edge Compute Plans

　　AT&T Labs has been quietly defining its concept of edge computers and is now slowly edging toward deploying them. Long term, the work has broad implications for the future design of both cloud and mobile systems.

　　AT&T defines edge compute as groups of servers and storage systems placed on the periphery of its network to deliver low latency services. It foresees a wide variety of such systems that vary in size and location depending on the application and demand.

　　“Edge compute is the next step in getting more out of our network, and we are busy putting together an edge computing architecture,” said Alicia Abella, a senior executive at AT&T Labs, in a keynote at the Fog World Congress here.

　　“We want to deploy edge compute nodes in mobile data centers, in buildings, at customers’ locations and in our central offices. Where it is…depends on where there is demand, where we have spectrum, we are developing methods for optimizing the locations,” she said.

　　The edge systems serve many uses. They aim to help AT&T reduce the volume of data it must carry to and from its core network. They also will enable higher quality for existing services and hopefully open the doors to new services as well.

　　One clear application is running video analytics for surveillance cameras. Such edge systems might use GPUs, FPGAs or other accelerators and be located in cities.

　　A more challenging use is handling jobs for automated vehicles because it would require a significant investment in roadside infrastructure but have an uncertain return-on-investment. Interestingly, AT&T now has 12 million smart cars on its network, a number growing by a million per quarter, she said.

　　Different edge systems will connect to AT&T’s core net in different ways. Some may take the form of one or two computing racks, while others could be larger.

　　One hardware difficulty the company has encountered so far is in getting small- and medium-sized versions of GPU accelerators. The products often come in large-scale, power-hungry versions, she said.

　　An early AT&T prototype put cameras in cars controlled by Raspberry Pi boards, sending video to remote monitors. Edge systems “could be as small as that, but more realistically we foresee some as small as a walk-in closet …we are only now starting to think about our strategy for deploying this,” she said.

　　Once such systems are widely deployed, they could have significant impact on mobile systems. Devices could offload work to the edge to save battery life.

　　“Imagine a new class of devices, or AR/VR without tethering to a computer but using mobile-edge compute…imagine a path to access signal processing in the edge,” she said.

　　Ultimately, AT&T aims to define a software platform for edge computing so developers can write applications for it. While the architecture is being created with that concept in mind, it will not likely be implemented until 2020 or later.

　　Abella speculated that in the distant future such a platform could enable crowdsourcing of compute and storage similar to the SETI@home project. Participants could even be compensated in bitcoin, she suggested.

　　The keynote was part of the Fog World Congress, where participants debated different definitions of edge, fog and cloudlet computing. On the show floor, some vendors displayed small, rugged gateways and industrial servers meant to serve such edge computing deployments.