Professor Shi Wei Matsumoto blog content from the 2016 review article published in the IEEE INTERNET OF THINGS JOURNAL, mainly summarizes the main content of the paper.
I. Introduction
Cloud computing is the last decade the most widely used way to provide computing services, is currently widely used technology companies mainstream SaaS software as a service business model is cloud computing. With the development of things, more and more devices access the network work together, some of the Internet of Things applications may require a very short response time, some of which may involve personal data, some of which may produce large amounts of data, which may be heavy network load. Cloud Computing In this scenario seems less suitable. Here to discuss the necessity of the second portion edge computing, and gives the definition edge computing. The third part introduces case studies of some edge computing. Section IV presents the possible challenges and opportunities. The fifth part conclusion to the challenges and outlook edge computing vision.
Second, what is the edge computing
1, pushing the limits due to the cloud: edge calculating means enabling technology for upstream data allow the calculation of the edge of the network, downstream data representing cloud services and networking service representatives. Here, we will "edge" is defined as any computing and network resources on the path between the data source and the cloud data center. In other words, our former C / S mode, not calculated between the client and server, all calculations are performed in the cloud. This is understandable, after all, cloud computing power is far stronger than the edge of the end. But as more devices to access the network development of new technologies, such as automatic driving, there are many devices will quickly generate a huge amount of data, such as sensor on the Boeing 787 aircraft will have 5GB of data per second, autonomous vehicles will also have 1GB of data per second, such a large amount of data to an existing wireless network technology is not enough to support real-time communication and cloud devices. And now, we will calculate the part placed near the edge of place.
2, the development of things for the lead: almost all electronic devices will be part of things, they will play the role of producers and consumers of data, such as air quality sensors, LEDbars, street lights, and even networked microwave. It is safe to infer, on the edge of the network, the number of things will grow to billions more in a few years. Thus, they produce the original data will be great, so that the traditional cloud sufficient to address all of these data. This means that most of the data generated by the Internet of Things will never be transmitted to the cloud, instead they will be processed at the edges.
3, consumer data becomes Manufacturer: In conventional cloud computing model, data consumers and producers are separated, in the edge of the network terminal equipment, such as smart phones, etc. In many scenarios, as is often consumer data, as shown in a. However, sometimes also as producer data terminal equipment, such as people to upload their own Vlog just took to the B station, circle of friends to share photos. In this scenario, video, photos, etc. previously uploaded to the cloud, the compression processing may be performed first at the local, network bandwidth in order to reduce pressure.
Third, the edge computing Case
1, cloud uninstall: at the edge of the calculation, the edge has a certain amount of computing resources, which provides an opportunity to extract part of the work load from the cloud. In a conventional content delivery network, only the data is stored on the edge server. This is based on the fact that the data provided by the supplier on the internet, this is the last few decades. In the IOT, production and consumption data at the edge. Thus, the edge computing paradigm, not only the data, and the operation applied to the data should be cached on the edge.
2, video analysis: spread of mobile phones and network video camera such analysis becomes an emerging technology. Due to long data transmission delay and privacy issues, cloud computing is no longer required for the application of video analytics. Here we cite looking for a lost child in the city an example. Today, a wide range of different types of cameras deployed in the urban areas and per vehicle. When a missing child, the child is likely to be captured by the camera. However, due to privacy issues or transportation costs, data in the camera usually not be uploaded to the cloud, which makes it extremely difficult to make use of wide-area camera data. Even if the data can be accessed in the cloud, upload and search large amounts of data can take a long time, this search for missing children is intolerable. Using edge computing paradigm, the cloud may be generated from the child search request, and push them to the target area of all things. For example, a smartphone can perform the request and searches its local camera data, only report the results to the cloud. In this mode, data may be utilized and computing power of all things, and the calculated result is obtained more quickly than the individual clouds.
3, Smart Home: Taking into account the data transfer pressure and privacy, the family's private data should be used primarily in the home. This feature makes cloud computing paradigm is not suitable for smart home. However, the edge computing is considered to be the perfect smart home form: can be easily connected to a Border Gateway running specializededge operating system (edgeOS) at home, things at home and management, data can be processed release the burden of Internet bandwidth locally, and services can also be deployed in edgeOS would be better
4, Smart City: calculated based on data in the local edge computing features locally produced and used, can greatly reduce the need for large urban data WAN, reduce latency, reduce the demand for data centers.
5, the edge of Cooperation: We can say that the cloud has become the de facto academia and industry to handle large data computing platforms. An important commitment behind cloud computing is that data should already exist or are being transferred to the cloud, and will eventually be processed in the cloud. However, in many cases, because of the huge costs privacy issues and data transmission, stakeholders have very little to share data with each other. Therefore, the opportunities for cooperation between the parties of shareholders is limited. Edge small on one physical data center, it is connected to the end user and clouds, with data processing capabilities, it can also be part of the logical concept. Edge theory proposed collaboration, connecting the edges of different geographic locations, different network structures of stakeholders. These are similar to ad hoc connections edge to the beneficiaries provided the opportunity to share data and collaborate. For example, cooperative medical and health issues.
Fourth, the challenges and opportunities
1, programmability: In the cloud computing environment, developers often deploy the project to the cloud, developers decided it was written by which programming language to use what technology, which is a device used for the user or program completely transparent, we only need to set the data interface; but at the edge of the computing environment, we need to deploy compute nodes to the edge, developers will face edge heterogeneous computing platforms, different operating environment, different communication protocol, different runtime. To solve these problems, researchers have proposed the concept of flow calculation , a set of functions is defined as the propagation path along the data applied to the data / calculation. I.e., along the propagation path data, as long as the application definition should be calculated where calculation can occur anywhere on the path. In this case, the flow computing can help users determine what function should be executed / calculated, and how to disseminate the data after edge computation.
In the process, it creates optimization problem. In fact Lynx double eleven concurrent processing, statistical data, the idea is to use the flow calculations. In this process, we need to deal with a delayed drive, energy restrictions, TCO (total cost of ownership that is) there are certain restrictions such as hardware and software, which requires researchers to design reasonable calculation unload way to the optimal solution offloaded to the calculated edge node. At the same time, developers will face calculate synchronization issues, migration and data migration operations, which are required to co-ordinate multiple layers edge computing collaboration.
2，命名 Naming：在边缘计算中，一个重要的假设是事物的数量非常大。在边缘节点的顶部，有大量的应用程序在运行，每个应用程序都有自己的服务提供结构。与所有计算机系统相似，边缘计算中的命名方案对于编程、寻址、事物识别和数据通信都是非常重要的。然而，边缘计算范式的有效命名机制尚未建立和标准化。边缘计算的实施者通常需要学习各种通信和网络协议，以便与系统中各种不同的设备进行通信。边缘计算的命名方案需要处理事物的移动性、高度动态的网络拓扑、隐私和安全保护，以及针对大量未信任的设备的可扩展性。
传统的命名方式如DNS的域名/IP映射、URI，很好的解决了互联网的命名问题。但这些方式不适用于边缘计算的移动性，因为有时大多数东西的边缘可能是高度移动和资源有限的。此外，对于网络边缘的一些资源受限的情况，考虑到其复杂性和开销，基于IP的命名方案可能过于繁重而难以支持。NDN和MobileFirst等命名规则可以适用于边缘计算场景中，但它们都有各自的缺点：NDN为以内容/数据为中心的网络提供了一个层次结构的名称，它对服务管理是人性化的，并提供了良好的可扩展性。但是，它需要额外的代理才能加入其他通信协议，如蓝牙或ZigBee等。与NDN相关的另一个问题是安全性，因为很难将硬件信息与服务提供者隔离。MobileFirst可以将名称与网络地址分离，以提供更好的移动性支持，如果将其应用于具有高度移动性的边缘服务，将非常有效。但是，需要使用全局唯一身份认证(GUID)来命名的是MobileFirst，而这在网络边缘的相关固定信息聚合服务(如家庭环境)中是不需要的。MobileFirst对于edge的另一个缺点是服务管理困难，因为GUID对人不友好。
对于如智能家庭环境的小型边缘计算场景，我们可以采取让edgeOS为每个事物分配网络地址的解决方案。在一个系统中，每个东西都有一个独特的人类友好的名字，它描述了以下信息:位置(在哪里)，角色(谁)，以及数据描述(什么)，例如，“厨房。烤箱。温度。”对于用户和服务提供者，这种命名机制使管理变得非常简单。例如，用户将收到来自geos的消息，如“灯泡3(什么)的天花板灯(谁)在客厅(哪里)坏了”，然后用户可以直接替换坏掉的灯泡，而不需要搜索错误代码或重新确定新灯泡的网络地址。此外，这种命名机制为服务提供者提供了更好的可编程性。同时，它阻止服务提供商获取硬件信息，这将更好地保护数据隐私和安全。但在如智慧城市这样的广阔的边缘计算场景中，这种方法并不适用，其命名机制还需要研究人员进行大量的工作。
3，数据抽象：有了物联网，网络中会有大量的数据生成器，这里我们以智能家居环境为例。在智能家居中，几乎所有的东西都会向edgeOS报告数据，更不用说部署了大量的东西。然而，大多数事情在网络的边缘，只有定期报告感知数据的网关。例如，温度计可以每分钟报告一次温度，但是这些数据很可能一天只被真正的用户使用几次。另一个例子可能是家里的一个安全摄像头，它可能会持续记录并将视频发送到网关，但数据只会存储在数据库中一段时间，没有人真正使用它，最终会被最新的视频覆盖掉。基于这一观察，我们设想人类对边缘计算的参与应该最小化，边缘节点应该消费/处理所有的数据，并以一种主动的方式与用户交互。在这种情况下，应该在网关级预先处理数据，如噪声/低质量的去除、事件检测和隐私保护等。进程数据将被发送到上层以提供未来的服务。在这个过程中会有几个挑战。
首先，来自不同事物的数据以不同的格式报告；出于对隐私和安全的考虑，在网关上运行的应用程序应该对原始数据保密。他们应该从整合的数据表中提取他们感兴趣的知识。其次，有时很难确定数据抽象的程度。如果太多的原始数据被过滤掉，一些应用程序或服务就不能学到足够的知识。然而，如果我们想要保留大量的原始数据，数据存储就会面临挑战。数据抽象的另一个问题是对事物的适用操作。收集数据是为了服务于应用程序，应用程序应该有控制(例如，从和写)权限，以完成某些服务用户的愿望。结合数据表示和操作，数据抽象层将作为连接到edgeOS的所有事物的公共接口。此外，由于事物的异构性，数据表示和允许的操作可能会有很大的差异，这也增加了通用数据抽象的障碍。
4，服务管理：在网络边缘的服务管理方面，我们认为应该支持以下四个基本特性来保证可靠的系统，包括differentiation, extensibility, isolation, 和 reliability.翻译过来就是差异性，可扩展性，独立性和可靠性。
差异性：随着物联网部署的快速增长，我们预计多种服务将部署在网络的边缘，如智能家居。这些服务将有不同的优先级。例如，关键服务，如故障诊断和报警，应该比普通服务处理得更早。例如，与娱乐等服务相比，降检测或心力衰竭检测也应该具有更高的优先级。
可扩展性：在网络的边缘，可扩展性可能是一个巨大的挑战，不像移动系统，设备等在物联网可能是非常动态的。当主人购买了一件新东西，它可以很容易地添加到当前的服务没有任何问题吗?或者当一个东西由于磨损而被替换时，以前的服务可以很容易地采用一个新节点替换吗?这些问题的解决需要在edgeOS中采用灵活的、可扩展的服务管理层设计。
独立性:隔离是网络边缘的另一个问题。在移动操作系统中，如果应用程序失败或崩溃，整个系统通常会崩溃并重新启动。或者在分布式系统中，可以使用不同的同步机制(如锁或令牌环)来管理共享资源。然而，在智能edgeOS中，这个问题可能更加复杂。可能有几个应用程序共享相同的数据资源，例如，光的控制。如果一个应用程序失败或没有响应，用户应该仍然能够控制他们的灯，而不会破坏整个系统。
可靠性:最后，可靠性也是网络边缘的一个关键挑战。我们从服务、系统和数据的不同角度来确定可靠性方面的挑战。A.系统应当在有设备损坏时及时提供给卡湖信息，甚至是在设备可能失效前就让用户有预感。B.系统中的每个组件都能够向edgeOS发送状态/诊断信息。有了这个特性，故障检测、事物替换和数据质量检测等服务可以很容易地部署在系统级。C.从数据的角度看，可靠性的挑战主要来自于数据传感和通信部分。正如之前所研究和讨论的，网络边缘的事物可能会因为各种原因而失败，它们也可能在不可靠的情况下报告低保真度的数据。多种新的物联网数据采集通信协议已经被提出，这些协议可以很好地支持大量的传感器节点和动态的网络状态。然而，连接的可靠性不如蓝牙或WiFi。如果传感数据和通信都不可靠，如何利用多个参考数据源和历史数据记录来提供可靠的服务仍然是一个开放的挑战。
5，保密性与安全性：在网络的边缘，使用隐私和数据安全保护是最重要的服务。如果一个家庭使用物联网，就会有很多隐私信息的产生，比如可以从感知到的使用数据中获取信息。例如，通过电或水的使用情况，我们可以很容易地推测房子是否空着。在这种情况下，如何保证隐私数据的安全，就是一个严峻的挑战。
6，优化指标：在边缘计算中，我们有多个具有不同计算能力的层。工作负载分配成为一个大问题。我们需要决定哪个层来处理工作负载，或者在每个部分分配多少任务。有多种分配策略来完成一个工作负载，例如，在每一层均匀分配工作负载，或者在每一层尽可能多地完成工作负载。极端情况是完全在端点上操作或完全在云上操作。为了选择最优分配策略，我们下面讨论几个优化指标，包括延迟、带宽、能源和成本。
延迟:延迟是评估性能的最重要指标之一，特别是在交互应用程序/服务中。云计算中的服务器提供了高计算能力。他们可以在相对较短的时间内处理复杂的工作，如图像处理、语音识别等。然而，延迟不仅由计算时间决定。长时延直接影响实时/交互密集型应用程序的性能。为了减少延迟，工作负载最好是在最近的一层，它有足够的计算能力到网络的边缘。
带宽:从延迟的角度来看，高带宽可以减少传输时间，特别是对于大数据(如视频等)。对于短距离传输，我们可以建立高带宽无线接入，将数据发送到边缘。一方面，如果工作负载可以在边缘处处理，那么与在云中工作相比，延迟可以大大改善。边缘和云之间的带宽也得到了节省。
能量:对于处于网络边缘的事物来说，电池是最宝贵的资源。对于端点层，将工作负载加载到边缘可以看作是一种能量免费的方法。那么，对于给定的工作负载，将整个工作负载(或部分工作负载)加载到edgerather而不是在本地计算，是否更有能量?关键是计算能耗和传输能耗之间的权衡。。一般来说，我们首先需要考虑工作量的功率特性。计算强度大吗?它将使用多少资源来运行本地?此外网络信号强度、数据大小和可用带宽也会影响传输能量开销。当传输开销小于局部计算时，我们倾向于使用边缘计算。然而，如果我们关注整个边缘计算过程而不是只关注端点，那么总能耗应该是每个使用层的能耗成本的累加。与端点层相似，每一层的能量消耗可以估计为局部计算成本加上传输成本。在这种情况下，最优的工作负载分配策略可能会改变。例如，本地数据中心层很忙，因此工作负载不断上传到上层。与端点计算相比，多跳传输可能会显著增加开销，从而导致更多的能量消耗。
成本:从服务提供商的角度，如YouTube、Amazon等，边缘计算为他们提供更少的延迟和能源消耗，潜在的增加吞吐量和改进的用户体验。因此，他们可以赚更多的钱来处理相同单位的工作量。例如，根据大多数居民的兴趣，我们可以把一个流行的视频在建筑层边缘。城市的层边缘可以摆脱这个任务，从而释放算力处理更复杂的工作。总吞吐量可以增加。服务提供商的投资是构建和维护每一层的成本。
工作负载分配不是一项简单的任务。这些度量标准是互相紧密相关的。例如，由于能源限制，工作负载需要在城市层上完成。与构建服务器层相比，能量限制不可避免地影响延迟。应该为不同的工作负载赋予指标优先级(或权重)，以便选择合理的分配策略。此外，成本分析需要在运行时进行。还应该考虑并发工作负载的干扰和资源使用。
五、总结
现在，越来越多的服务被从云端推到网络的边缘，因为在边缘处理数据可以确保更短的响应时间和更好的可靠性。此外，如果更多的数据可以在边缘处理而不是上传到云端，带宽也可以节省。物联网的蓬勃发展和移动设备的普及改变了edge在计算范式中的角色，从数据消费者转变为数据生产者/消费者。在网络的边缘处理或修改数据会更有效。传统的云计算模式仍然受到支持，但由于数据的紧密性，它还可以将远程网络连接在一起进行数据共享和协作。最后，提出了在可编程性、命名、数据抽象、服务管理、隐私和安全以及优化指标等方面值得研究的挑战和机遇。边缘计算就在这里，我们希望这篇论文能引起社会的关注。