Bringing Intelligence to the Network Edge in Order to Address Heightening Industrial Demands

5 min read

The Winter Olympics held in Korea at the start of this year was not just about sports on snow and ice. The event showcased the first wireless network to take advantage of the dramatic improvements in performance offered by the emerging international 5G mobile communication standard.

A self-driving bus took spectators around the Alpensia resort which hosted a number of the competitions and relied on 5G to help it navigate. Inside these vehicles, it used the high-bandwidth wireless link to relay a live 3D television feed to screens so as to entertain passengers. On the cross-country course, skiers wore GNSS receivers that let servers know where they are at any point in the race. Computers employed data taken from surveys of the course to generate live views of what the skiers could see ahead of them. The video was then piped to attendees who had experimental 5G handsets.

Although the Korea Telecom trials focused on the applications of 5G for entertainment purposes, the wireless network standard will be fundamental to a new crop of industrial applications. The demonstrations at the 2018 Winter Olympics focused heavily on showing off large improvements in bandwidth over 4G. The objective for 3GPP developing the 5G standard, however, is to take mobile communications beyond its current parameters and attend to new prospective openings in industrial automation and IoT. Though bandwidth, as well as the far lower power requirements it has for each data connection, will be important, it is another factor that is set to drive 5G’s adoption in real-time industrial applications - such as process monitoring, factory automation systems, control of robots, fleet management and suchlike. That factor is latency.

Existing wireless networks exhibit round-trip latencies in the order of hundreds of milliseconds. To improve aggregate data delivery rates, messages from multiple handsets and other roving devices are interleaved and processed together. 5G changes this policy to make it possible to respond to incoming messages in as little as a single millisecond. Reducing the latency to such a level means that a deterministic approach can be taken and thereby opens up new possibilities in the industrial arena. It will thus now be feasible to bring distributed processing to real-time mission-critical systems that once demanded their compute resources to be entirely local.

One of the fundamental problems of immersive virtual reality (VR) is the lag between the user moving their head and the updated view they can see in their headset. Typically, people can deal with a ‘motion-to-photon’ latency of up to 20ms. Any more and they quickly start to suffer from the effects of motion sickness. One benefit of offloading VR processing to remote servers is that it greatly reduces the power consumption of the body-worn electronics and the uncomfortable heat that circuits tend to generate. By bringing latencies well under the motion-to-photon barrier, VR can benefit greatly from 5G technology. These more advanced VR systems will then be able to support industrial applications, such as the remote control of robots needed to make repairs in hazardous locations where it isn’t safe for human members of the workforce to go.

Robotic systems with greater freedom of movement suffer from a similar problem to virtual reality. The control algorithms rely on low delays to maintain stability. For the relatively simple single-function robots used in high-volume production today this is not a major problem. But robots are increasingly becoming ‘cobots’. They are no longer contained within caged work cells. They work closely with people and other robots, moving freely around the workplace. To do so, they require more sophisticated software, which includes the new wave of compute-intensive technologies based on deep learning.

Mobile robots will run on battery power, and they will not be isolated to the factory. As robots move into sectors such as agriculture and logistics, where they need to take to the outside world, their power consumption will be a major obstacle to bringing on board the sophisticated algorithms that they run. 5G wireless networks provide the ability to offload power-hungry software to remote servers that are powered by reliable mains electricity. The result of this division of responsibilities is a group of robots that can learn from shared experience without draining their batteries in a matter of minutes.

The servers, however, cannot be too remote. The speed of light passing through fibre-optic cables and switching delays incurred by every router along the path play a role in determining network responsiveness. The rule of thumb is that every 100km of distance adds a millisecond of delay in each direction. To reap the benefits of 5G’s low latency, autonomous systems will need to deploy server resources closer to the point of demand than is typical today. It is this, of course, that is leading to the development of the edge computing market.

Edge computing servers may sit within an industrial campus. For example, a mining operator that makes extensive use of autonomous vehicles to move material around a site can choose to deploy their own micro data centre. In many cases they will have the ability to make use of shared, cloud-based services located a little further away to provide the edge computing resource. Although mines are frequently in remote locations, many are close enough to urban centres to make it viable for service providers to set up their own edge computing resources that are shared with other agricultural and industrial users.

Computing at the edge of the cloud will not replace the centralised servers that form the basis of today’s network architectures. There will still be a need for longer-term planning, which will be handled by a second layer of software with machine learning at its core. The portions that need a real-time response, however, will be devolved to the ‘cloudlets’ that lie between wireless nodes and the core cloud servers.

Cloudlet computing requires a change in approach to server design from what is conventional in core data centres. Performance remains critical, but it has to be compact, reliable and energy-efficient too. The design of the compute elements is likely to have more in common with that of cellular base stations than high-end server blades. Space, cooling and opportunities for maintenance will be limited. Some smaller cloudlets may coexist with telecom equipment in roadside cabinets, while others will be fitted into containers for easy and rapid deployment.

Architectures designed for cloudlets will make extensive use of acceleration technologies for tasks such as machine learning. Accelerators offer higher performance per Watt than general-purpose multicore processors. Operational reliability can be delivered through the use of redundant nodes, but a key challenge lies in security. Physical security is harder to guarantee for cloudlets compared to large-scale data centres. Tamper and intrusion detection will be needed so that if attackers gain access to the hardware, systems can take action and move workloads to unaffected cloudlets nearby.

Computer design will play a key role in guaranteeing security against attacks launched over the network. Very different applications will need to run alongside each other on the cloudlet hardware to provide high resource utilisation. The cloudlet systems need to be able to ensure that applications running on one node cannot interfere with others and, in doing so, intercept private data or disrupt operations. Technologies such as virtualisation and containerised software will make it easier to maintain application separation, monitor behaviour and let administrators lock out bad actors. Software running in containers will also be easier to move as resource requirements change, thereby helping cloudlet operators optimise their infrastructure and guarantee performance levels.

The arrival of 5G, with its markedly reduced power consumption per data connection and deterministic capabilities, will clearly be highly advantageous to the industrial sector. Furthermore, it represents the beginning of a pivotal stage in the ongoing evolution of cloud computing. Through improvements in hardware and software technology, cloudlets located close to the point of need will make it possible to realise a new generation of artificial intelligence enabled real-time systems and, in doing so, drive forward the next industrial revolution.