With the arrival of 5G, the edge computing will become one of the key innovations to a new era of change communication information service mode. There mechanism predicts that 50% of the 2025 data will be analyzed at the network edge side, processing and storage, but also is considered to be edge computation and industrial 5G Internet, networking and other important binding site, can bring more subversive business model.
But with further research and practice, we found that edge will not calculate compute nodes (such as servers, storage devices, etc.) to move into the edge of the network room can sit back and relax, in fact, can really achieve the desired edge computing needs of the entire information infrastructure upgrade facilities and evolution, in other words, the edge computing network reconfiguration needs 5G + + + cloud computing room renovation + ...... This article from the beginning edge computing features, focuses on the important relationship between the edge computing and network reconfiguration, edge calculation of network edge computing needs and the impact on the network might bring, as well as some of the current research.
Edge computing needs for network
First, low latency characteristics of edge computing is one of the most important features, but also different from the traditional key element of cloud computing, in other words, reach a stable low-latency metrics program can not be called edge computing. But to achieve low latency, is not simply down to the compute nodes can be solved, needs to consider the bearer network level, the level of light transmission level or even a cable network layout, Figure 1 is a case observed in practice , an industrial campus network deployment plans edge computation, but here the position at the junction of the two administrative areas, although the base station and the edge of the test client access compute nodes deployed close to a straight line distance, but with the device, but in both a different access ring, the client traffic to the edge node to calculate the actual flow through the bypass is required to deploy to a convergence layer and then another even core devices, cause the actual measured delay is much higher than expected.
The other scenario is a typical wide attention and high expectations of car networking scene, the car requires extremely low latency network, but the network is not along the road network, then the situation can easily position the distance from the actual network appears far Therefore, while the deployment of edge computing, based on business needs features to re-plan design of the network architecture.
FIG 1 unoptimized bearer network edge computing affect implementation effect
Secondly, the internal edge node calculation scheme will affect the calculated edge index low latency. For example (Figure 2), AI AI inference part applications require low latency HPC, consider deploying edge node calculation, and the AI model training is part of a large number of non-real-time calculation, consider deploying cost large centralized relatively low cloud computing platform. However, the existing calculation programs generally on edge by a traditional data center networking solutions or cloud computing platform, a two-layer multi-switch common program (convergence + Spine-Leaf access architecture or architecture) + outlet layer networking mode router.
But in the relevant test verification, we found a new problem: First, such a network architecture virtually on the edge of the internal computing nodes increases the number of routes and device nodes, increasing the delay, and the delay is not controllable, expected indicators are quite different; the second is such a traditional data center networking solutions is a lossy network, in a play more extreme cases, namely when the overall network load is not high, there will be congestion and packet loss will result in AI computing performance dropped significantly, three had sunk to the edge for a base station or access room calculation, it may be a server rack can be placed, also in the number of servers 10 to 20, in this case, still network architecture using the traditional data center model, apparently in investment earnings ratio is relatively poor, not easy operation and maintenance.
2 AI node network diagram showing a calculation application edge
Impact on the network edge computing
On the other hand, the edge will bring a lot of computing new traffic bearer network, but the traffic on the conventional flow characteristics quite different. In the above-described Example AI cases, large amounts of data need to be deployed on the other side of the cloud computing node network, but this process no transmission quality requirements, a non-real-time traffic class, may also want to reduce costs.
Apparently adopted to calculate the edge line between the node and the bearer node is a cloud computing program extremely uneconomical, so more we tend to divert the flow through this part of the common Internet. But this is in fact what traffic flows between data centers with traditional Internet traffic (north-south traffic) there is a big difference in the traffic model. If you adhere to the traditional network architecture unchanged, easily lead to a lot of traffic roundabout, affecting the normal transmission of other services. Bearer network is required to rebuild a flexible architecture that can meet the traditional north-south traffic, but also relatively flexible to adapt to new things flow.
The introduction of new technology is the future research directions
因此随着边缘计算的出现与部署，承载网层面需要应对新出现的业务需求与流量特征，对网络架构进行重构，将传统以承载南北向流量为主的基础网络架构，向能够灵活调度、兼顾时延指标、利于东西向流量的新型融合网络架构的方向发展。
目前业界已经注意到这些问题，多种新的思路或方案已经被提出，并在不同层面进行验证。典型方案可以分为有两个层面：一是网络架构层面，采用新型的网络架构来替代传统的承载网络，比如新型城域网架构等；二是采用新的技术体系来解决传统IP体系无法解决的问题，比如无损网络技术等。
在网络架构层面，要满足边缘计算带来的新需求，即满足边缘计算对低时延特性的需求，也要能够应对边缘计算带来的新特征的流量。因此构建新型城域网架构成为新的方向，新型城域网架构以采用通用设备组网为主，将传统的树形网络架构演进到基于Spine-Leaf架构，实现固定和移动网络的融合统一承载，同时引入FlexE、SR、EVPN等技术，提供差异化服务能力，为不同客户群提供不同等级的切片网络。
在网络技术层面，针对边缘计算节点内部网络技术，需要分为高性能计算需求场景和小规模计算节点场景两个不同类型的场景分别进行讨论。针对高性能计算需求场景，虽然传统的InfiniBand技术能够解决问题，但它与其他网络技术兼容较小，需要专门设备，如果全网性的规模部署会对运营维护造成较大压力。因此一种可行的解决方案是采用兼容主流IP与以太网协议体系的无损网络体系，如以RoCEv2为主等RDMA方案等。目前已经有主流设备厂家提供数据中心交换机的升级版本来支持RoCEv2等协议族，IETF等主流标准组织也开展了相应的标准化进程。
Under the scenario for small computing nodes, such as access to the room after the transformation, which is limited physical space, power, and air conditioning capacity limited circumstances, be less able to put the servers and storage devices, may be less to a chassis size, but also can not be a traditional 1100mm ~ 1200mm deep server cabinet, only 600mm deep, 300mm or even communication chassis, for which either the server level or at the network level, the device will need to customize new device form. Such as the Open Data Center Committee (ODCC) initiated OTII (Open Telecom IT Infrastructure) project launched using telecommunications equipment standards, depth of less than 600mm, can be mixed OTII server deployment and telecommunications equipment to meet the edge of the compute nodes of special requirements . Currently three operators are organizing an open research network convergence device, through a modular way, on-demand configuration of the network, computing and storage capacity, the ability to avoid wasting equipment and floor space.
To sum up, the edge computing will become an important new era 5G business model will bring new opportunities and challenges in the network, so the industry needs in the next step in the evolution of the network for the network edge computing needs and edge calculate the effect on the network to bring two angles, re-examine the idea of network development, the introduction of new forms of technology and equipment to build a new generation of information infrastructure.
