Author | Yudong Huang, Ph.D. 1, State Key Laboratory of Networks and Exchanges, Beijing University of Posts and Telecommunications, research direction is network architecture, deterministic networks, email: hyduni@163.com.

At present, deterministic networks have become a global research hotspot and urgently needed key direction. On the one hand, under the background of the grand strategy of the industrial Internet, industries such as steel, petroleum, mining, ports, logistics, and manufacturing are deeply integrated with ICT technology, access equipment is growing massively, and the traditional bus network is gradually replaced by standardized deterministic Ethernet, which will help the industry upgrade to intelligent and digital transformation, and achieve quality reduction, cost reduction and efficiency increase. On the other hand, emerging time-sensitive scenarios and applications, such as the Internet of Vehicles, metaverse, holographic communication, remote remote operation, remote diagnosis and treatment, remote driving, remote equipment maintenance, immersive real-time tourism, interactive distance education, online interactive concerts, and tactile Internet, have stimulated people's imagination of the future Internet. And bring a trillion-level blue ocean market.

Taking metaverse applications as an example, from the perspective of technological development history, the Internet is evolving from an app application that transmits information to a metaverse that is deeply integrated with production, life, health and entertainment. From 2010 to 2020, with the upgrading of infrastructure such as mobile Internet and cloud computing, national applications such as Douyin, WeChat, and Taobao were born. Since 2020, a new round of industrial technological change has begun, Facebook has changed its name to Meta, Google Map has launched immersive reality 3D maps, Microsoft has acquired game giant Activision Blizzard, Byte has acquired VR startup Pico, Tencent has made efforts to XR True Internet, and domestic and foreign giants have laid out to seize the commanding heights of the "meta-universe". In addition, for hardware manufacturers such as Apple and Xiaomi, global mobile phone shipments fell by 11% in the first quarter of 2022, and domestic mobile phone market shipments in March fell by 40.5% year-on-year, and smart wearable devices such as headsets, glasses, headphones, watches, bracelets, and handles that can bring a new interactive experience or subvert mobile phone screens have become new growth points.

In the next decade, whether the network infrastructure can achieve massive real-time interoperation of "human-machine-thing" will be the key to whether China can take the lead in nurturing the metaverse. However, as the core link of the Internet, it does not yet have deterministic service quality assurance capabilities such as delay, jitter, and packet loss rate, and uncertain delay will lead to immersive experience lag, interaction dizziness, user dissuasion, and even possible security accidents. Therefore, the application must not arrive, the network first.

The network itself is a complex system full of uncertainty, including terminal equipment, network equipment, and computing and storage resources; Horizontally, the network passes through multiple autonomous domains, such as local area networks, access networks, aggregation networks, backbone networks, and cloud data centers. Longitudinally, the network covers 5G, WiFi, optical, Ethernet, TCP/IP, Quic, http and other exquisite protocol technologies at different levels from hardware boards to software operating systems. The Internet has been tempered for more than 50 years, and "best effort" forwarding is its foundation, requiring it to have certainty is a bit square and round, like looking good and wanting to buy new shoes are always not fit.

In order to allow the traditional network to smoothly evolve to a deterministic network, and to facilitate discussion and research, this paper proposes three possible principles for everyone to define deterministic networks.

The premise of deterministic networking is deterministic traffic

If we can strictly guarantee the service quality of the application, it must be because we know the traffic of the application enough. Computers and mobile phones generate user traffic, generally their traffic arrival curve can be regarded as Poisson distribution, that is, when the user uses the traffic continues to send for a period of time, when not in use, the curve becomes a normal distribution, so it has a long tail delay; Traditional networks are not aware of applications, and can only optimize bandwidth and reduce average latency. In industrial scenarios, sensors, controllers, and actuators generate machine traffic, and their traffic characteristics are very clear, such as control instructions sending a 100-byte packet every 1ms, occupying 0.8Mbps bandwidth, and its arrival curve is an average distribution without bursts.

In addition, the slow start and fast retransmission of computer TCP congestion control will bring a large number of traffic bursts, while many industrial equipment (such as PLCs) do not have TCP/IP protocol stacks, which are directly packaged into frames in the second-layer MAC and sent out. Taking the on-board network as an example, most of its deterministic traffic in addition to audio and video entertainment is in line with this principle, so it is completely feasible to replace the messy bus in the car with Ethernet-based time-sensitive network technology to reduce the weight of the body, increase the body space, and improve the level of automotive intelligence.

So if the traffic is a little more complex and not completely certain, is a deterministic network still possible? Yes. At this time, we need to make the network aware of the application, sign an SLA service level agreement with the application, such as promising maximum bandwidth and burstiness, and then reserve network resources such as bandwidth, time slots, and cache along the path according to demand. Therefore, deterministic networks are usually also related to global control, access control, connection establishment, resource reservation, network calculus and other technologies. If the traffic does not come as promised, the oversent traffic may be dropped, or the promised traffic arrival model can be met by various traffic shaping methods, of course, shaping itself will also introduce additional delay. Furthermore, if the traffic is completely arbitrary, then a deterministic network is impossible.

Therefore, "deterministic" is equivalent to adding constraints, the more constraints, the more certain, and "deterministic traffic" is the largest premise constraint in the "deterministic network".

Deterministic traffic is only a small fraction

Suppose we build a deterministic network with a port rate of 100G, can we carry 100G deterministic traffic? It's almost impossible. And if the deterministic traffic only accounts for 10G, and the remaining 90G is still best-effort traffic, this is more achievable. Next, the canteen queuing scene and the ambulance scene are used to analyze why deterministic flow is best only a small part.

In the cafeteria queuing scenario, if a company has 1,000 employees who all leave work at 12 noon and go to the cafeteria to eat, those at the back of the line will inevitably face longer waiting times. Since everyone is equal, no matter how the scheduling is optimized and how the queue is cut, the last "deterministic employee" will not get a good delay jitter guarantee. Network resources are finite, and the determinism of the network and the reuse of resources are the poles of the magnet. For example, the leased line of pulling optical fiber has high certainty, but it can only transmit point-to-point, unable to route switching and multicast, and the resource multiplexing rate is low; Ethernet adopts statistical multiplexing and high resource utilization, but when the bandwidth utilization reaches 30%-50% or higher, a large amount of packet loss will be generated, and hard slicing, priority queues, etc. are required to isolate to reduce the resource reuse rate, thereby improving deterministic guarantee.

In the ambulance scenario, there is a special emergency lane on the highway, and ordinary vehicles will also take the initiative to give up road resources for the ambulance, so the ambulance gets the right of priority transmission and can reach the destination within the determined delay. In the case of many ambulances, we can also adjust the traffic lights and other ways to keep the ambulances at a distance from each other and isolate them in time and space. In addition, this also explains that deterministic networks will not replace the traditional Internet, because it is forward-compatible, and traditional best-effort traffic can still run on ordinary roads.

Thus, certainty is equivalent to priority, which predestines certainty to be only a small part.

Certainty, flexibility, and scalability can only be two

In the early technological evolution, there was the dispute between ATM and Ethernet, and ATM could strictly ensure the quality of service by establishing a connection to pass a similar circuit switching of cells, but it was not widely used. At the same time, the QoS technology of the Internet has two models, DiffServ and IntServ, of which IntServ can also provide deterministic service quality assurance for special applications through end-to-end RSVP resource reservation, but because of the need to maintain the state of each stream, the technology lacks scalability and cannot cope with massive traffic scenarios, so it has not been widely used. The result is Ethernet/IP with flexibility and scalability.

Since Ethernet has the advantages of large bandwidth, low cost and good compatibility, now we want to implement deterministic networks based on Ethernet, and there is always a trade-off. If certainty is required, flexibility and scalability cannot be combined. For example, within the scope of the local area network, we can build a dynamic and flexible deterministic network through some technologies, but it will be limited by factors such as network radius and hop number. If we want to scale to large networks, realize deterministic transmission in complex topologies and massive traffic scenarios, and need scalability, flexibility will be greatly reduced. If a solution has certainty, flexibility, and scalability, then we probably need to think about whether its technical complexity and cost are appropriate, as well as the possibility of commercial landing.