Edge Computing: Vision and Challenges文章对边缘计算目前存在的问

文件名称: Edge Computing: Vision and Challenges

所属分类: 平台管理

开发工具:

文件大小: 468kb

下载次数: 0

上传时间: 2019-04-06

提供者: stm32_ne********

下载 (468kb)

不能下载？报告错误

详细说明：文章对边缘计算目前存在的问题和解决方法提供了很好的描述，并给出了边缘计算的概念。对入门有很好的帮助。SHI et al.: EDGE COMPUTING: VISION AND CHALLENGES C. Edge computing Benefits view can be generated immediately upon the user request In edge computing we want to put the computing at the reaching the edge node. Of course the data at the edge ne proximity of data sources. This have several benefits com- should be synchronized with the cloud, however, this can pared to traditional cloud-based computing paradigm. Here we done in the background use several early results from the community to demonstrate Another issue involves the collaboration of multiple edges ne potential benefits. Researchers built a proof-of-concept When a user moves from one edge node to another. One simple platform to run face recognition application in [20], and the solution is to cache the data to all edges the user may reach response time is reduced from 900 to 169 ms by moving com- Then the synchronization issue between edge nodes rises up putation from cloud to the edge. Ha et al. [21] used cloudlets All these issues could become challenges for future investi to offload computing tasks for wearable cognitive assistance gation. At the bottom line, we can improve the interactive and the result shows that the improvement of response time is services quality by reducing the latency. Similar applications between 80 and 200ms. Moreover, the energy consumption also include the following could also be reduced by 30%-40% by cloudlet offload- D)Navigation applications can move the navigating or ing. clonecloud in [22] combine partitioning, migration with searching services to the edge for a local area, in which merging, and on-demand instantiation of partitioning between case only a few map blocks are involved mobile and the cloud, and their prototype could reduce 20x 2)Content filtering/aggregating could be done at the edge running time and energy for tested applications nodes to reduce the data volume to be transferred 3) Real-time applications such as vision-aid entertainment IIL. CASE STUDY ames, augmented reality, and connected health, could make fast responses by using edge node In this section, we give several case studies where edge Thus, by leveraging edge computing, the latency and con- computing could shine to further illustrate our vision of edge sequently the user experience for time-sensitive application computing could be improved significantly A. Cloud Offloading B. Video analytic In the cloud computing paradigm, most of the computa- tions happen in the cloud, which means data and requests The widespread of mobilephones and network cameras are processed in the centralized cloud. However, such a make video analytics an emerging technology. Cloud comput computing paradigm may suffer longer latency (e. g, long ing is no longer suitable for applications that requires video tail latency), which weakens the user experience. Numbers analytics due to the long data transmission latency and privacy of researches have addressed the cloud offloading in terms concerns. Here we give an example of finding a lost child of energy-performance tradeoff in a mobile-cloud environ in the city. Nowadays, different kinds of cameras are widel ment [22]-[26]. In edge computing, the edge has certain deployed in the urban area and in each vehicle. When a child computation resources and this provides a chance to offload is missing, it is very possible that this child can he captured by part of the workload from cloud a camera. However, the data from the camera will usually not In the traditional content delivery network, only the data is be uploaded to the cloud because of privacy issues or traffic which makes it extremely difficult to leverage the wide cached at the ed This is based on the fact that th content provider provides the data on the Internet, which is area camera data. even if the data is accessible on the cloud true for the past decades. In the loT, the data is produced and uploading and searching a huge quantity of data could take a long time, which is not tolerable for searching a missing child consumed at the edge. Thus, in the edge computing paradigm, With the edge computing paradigm, the request of searching not only data but also operations applied on the data should be cached at the edge a child can be generated from the cloud and pushed to all the One potential application that could benefit from edge things in a target area. Each thing, for example, a smart phone computing is online shopping services. A cuStomer may can perform the request and search its local camera data and manipulate the shopping cart frequently. By default, all these only report the result back to the cloud. In this paradigm, it is changes on h s/her shopping cart will be done in the cloud possible to leverage the data and computing power on every and then the new shopping cart view is updated on the cus thing and get the result much faster compared with solitary tomer's device. This process may take a long time depending cloud computing on network speed and the load level of servers. It could be even longer for mobile devices due to the relatively low band- C. Smart Home width of a mobile network. As shopping with mobile devices lo T would benefit the home environment a lot. Some prod- is becoming more and more popular, it is important to improve ucts have been developed and are available on the market such the user experience, especially latency related. In such a sce- as smart light, smart TV, and robot vacuum. However, just nario, if the shopping cart updating is offloaded from cloud adding a Wi-Fi module to the current electrical device and servers to edge nodes, the latency will be dramatically reduced. connecting it to the cloud is not enough for a smart home As we mentioned, the users' shopping cart data and related In a smart home environment, besides the connected device operations(e. g, add an item, update an item, delete an item) cheap wireless sensors and controllers should be deployed to both can be cached at the edge node. The new shopping cart room, pipe, and even floor and wall. These things would report 640 IEEE INTERNET OF THINGS JOURNAL. VOL 3. NO. 5 OCTOBER 2016 Hospital App n Insurance logistics Service Management (DEIR) Differentiation(C) Extensibility(E) Isolation(0) ReliabilityIR Collaborative Edge Naming Data Abstraction Pharmacy Pharmaceutical Corrrmunication Wi-Fi BluetoothEthernet Government Fig 3. Structure of edgeS in the smart home environment g. 4. Collaborative edge example: connected health an impressive amount of data and for the consideration of data transportation pressure and privacy protection, this data edge computing is also an appropriate paradigm since it could should be mostly consumed in the home. This feature makes save the data transmission time as well as simplify the network the cloud computing paradigm unsuitable for a smart home. structure Decision and diagnosis could be made as well as dis- Nevertheless, edge computing is considered perfect for build- tributed from the edge of the network, which is more efficient ing a smart home: with an edge gateway running a specialized compared with collecting information and making decision at edge operating system (edgeOS) in the home, the things can central cloud be connected and managed easily in the home, the data can 3) Location Awarenes.: For geographic-based applications be processed locally to release the burdens for Internet band- such as transportation and utility management, edge computin width, and the service can also be deployed on the edges exceed cloud computing due to the location awareness In edge for better management and delivery. More opportunities and computing, data could be collected and processed based on potential challenges are discussed in Section IV geographic location without being transported to cloud Fig. 3 shows the structure of a variant of edgeS in the smart home environment. EdgeS needs to collect data from E. Collaborative edge mobile devices and all kinds of things through multiple com Cloud, arguably, has become the de facto computing plat- munication methods such as Wi-Fi, Blue Tooth, ZigBee, or form for the big data processing by academia and industry. A a cellular network. Data from different sources needs to be key promise behind cloud computing is that the data should fused and massaged in the data abstraction laver. Detailed he already held or is being transmitted to the cloud and will description of this process will be discussed in Section IV-C On top of the data abstraction layer is the service manage eventually be processed in the cloud. In many cases, however, ment layer. Requirements including differentiation, exten the data owned by stakeholders is rarely shared to cach other due to privacy concerns and the formidable cost of data tranS- sibility, isolation, and reliability will be supported in this layer. In Section IV-D, this issue will be further addressed portation. Thus, the chance of collaboration among multiple The naming mechanism is required for all layers with dif- stake-holders is limited. Edge, as a physical small data center ferent requirements. Thus, we leave the naming module in that connects cloud and end user with data processing capabil- a cross-layer fashion. Challenges in naming are discussed ty, can also be part of the logical concept. collaborative edge which connects the edges of multiple stakeholders that are in section iv-B geographically distributed despite their physical location and network structure is proposed [15]. Those ad hoc-like con D. Smart City nected edges provide the opportunity for stakeholders to share The edge computing paradigm can be flexibly expanded and cooperate data from a single home to community, or even city scale. Edge One of the promising applications in the near future is computing claims that computing should happen as close as connected health, as shown in Fig 4. The demand of geograph- possible to the data source. With this design, a request could ically distributed data processing applications, i.e., healthcare, be generated from the top of the computing paradigm and requires data sharing and collaboration among enterprises in be actually processed at the edge. Edge computing could be multiple domains. To attack this challenge, collaborative edge an ideal platform for smart city considering the following can fuse geographically distributed data by creating virtual characteristics shared data views. The virtual shared data is exposed to end 1) Large data Quantity: A city populated by 1 million peo- users via a predefined service interface. An application will ple will produce 180 PB data per day by 2019191, contributed leverage this public interface to compose complex services Dy public safety, health, utility, and transports, etc. Building for end users. These public services are provided by partici- centralized cloud data centers to handle all of the data is unre- pants of collaborative edge, and the computation only occurs alistic because the traffic work load would be too heavy. In in the participant's data facility such that the data privacy and this case, edge computing could be an efficient solution by integrity can be ensured processing the data at the edge of the network To show the potential benefits of collaborative edge, we 2)Low Latency: For applications that require predictable use connected healthcare as a case study. We use a flu out and low latency such as health emergency or public safety, break as the beginning of our case study. The patients fow SHI et al.: EDGE COMPUTING: VISION AND CHALLENGES 641 to hospitals, and the electronic medical record(EMr) of the where the computing is conducted in a cloud. Users have zero patients will be updated. The hospital summarizes and shares or partial knowledge of how the application runs. This is one the information for this flu outbreak, such as the average cost, of the benefits of cloud computing that the infrastructure is the symptoms, and the population, etc. a patient theoretically transparent to the user. Usually, the program is written in one will follow the prescription to get the pills from a pharmacy. programing language and compiled for a certain target plat- One possibility is that a patient did not follow the therapy. form, since the program only runs in the cloud. However, in Then the hospital has to take the responsibility for rehospi- the edge computing, computation is ofFloaded from the cloud talization since it cannot get the proof that the patient did and the edge nodes are most likely heterogeneous platforms. In not take the pills. Now, via collaborative edge, the pharmacy this case, the runtime of these nodes differ from each other, and can provide the purchasing record of a patient to the hospital, the programmer faces huge difficulties to write an application which significantly facilitates healthcare accountability that may be deployed in the edge computing paradigm At the same time, the pharmacies retrieve the population To address the programmability of edge computing, we of the flu outbreak using the collaborative edge services pro- propose the concept of computing stream that is defined as vided by hospitals. An apparent benefit is that the pharmacies a serial of functions/computing applied on the data along have enough inventory to obtain much more profits. Behind the data propagation path. The functions/computing could the drug purchasing, the pharmacy can leverage data provided be entire or partial functionalities of an application, and by pharmaceutical companies and retrieve the locations, prices the computing can occur anywhere on the path as long as and inventories of all drug warehouses. It also sends a trans- the application defines where the computing should be con port price query request to the logistics companies. Then the ducted. The computing stream is software defined computing pharmacy can make an order plan by solving the total cost flow such that data can be processed in distributed and effi- optimization problem according to retrieved information. The cient fashion on data generating devices, edge nodes, and pharmaceutical companies also receive a bunch of flu drug the cloud environment. As defined in edge computing, a orders from pharmacies. At this point, a pharmaceutical com- lot of computing can be done at the edge instead of the pany can reschedule the production plan and rebalance the centric cloud. In this case, the computing stream can help inventories of the warehouses. Meanwhile, the centers for dis- the user to determine what functions/computing should be ease control and prevention, as our government representative done and how the data is propagated after the computing in our case, is monitoring the flu population increasing at wide happened at the edge. The function/computing distribution range areas, can consequently raise a flu alert to the people metric could be latency-driven, energy cost, TCO, and hard in the involved areas. Besides, further actions can be taken to ware/software specified limitations. The detailed cost model is prevent the spread of flu outbreak discussed in Section IV-F. By deploying a computing stream After the Au outbreak, the insurance companies have to pay we expect that data is computed as close as possible to the the bill for the patients based on the policy. The insurance data source, and the data transmission cost can be reduced companies can analyze the proportion of people who has the In a computing stream, the function can be reallocated, and fu during the outbreak. This proportion and the cost for flu the data and state along with the function should also be treatment are significant factors to adjust the policy price for reallocated. Moreover, the collaboration issues(e.g, synchro the next year. Furthermore, the insurance companies can also nization, data/state migration, etc. have to be addressed across provide a personalized healthcare policy based on their Emr multiple layers in the edge computing paradigm if the patient would like to share it Through this simple case, most of the participants can ben- efit from collaborative edge in terms of reducing operational B. Naming cost and improving profitability. However, some of them, like In edge computing, one important assumption is that the hospitals in our case, could be a pure contributor to the health- number of things is tremendously large. At the top of the care community since they are the major information collector edge nodes, there are a lot of applications running, and each in this community application has its own structure about how the service is pro vided. Similar to all computer systems, the naming scheme in IV. CHALLENGES AND OPPORTUNITIES edge computing is very important for programing, addressing, We have described five potential applications of edge com- things identification, and data communication. However, an puting in the last section. To realize the vision of edge efticient naming mechanism for the edge computing paradigm computing, we argue that the systems and network commu has not been built and standardized yet. Edge practitioners nity need to work together. In this section, we will further usually needs to learn various communication and network summarize these challenges in detail and bIng forward some protocols in order to communicate with the heterogeneous potential solutions and opportunities worth further research, things in their system. The naming scheme for edge computing including programmability, naming, data abstraction, service needs to handle the mobility of things, highly dy ynamic network management, privacy and security and optimization metrics. topology, privacy and security protection, as well as the scal- ability targeting the tremendously large amount of unreliable A. Programmability In cloud computing, users program their code and deploy Traditional naming mechanisms such as DNs and uniform them on the cloud. The cloud provider is in charge to decide resource identifier satisfy most of the current networks very IEEE INTERNET OF THINGS JOURNAL. VOL 3. NO. 5 OCTOBER 2016 Service management Name Replacement Programmability Identifi 00001> I Things management ID Time Data Address MAC address, IP addr i Communication protocol Fig. 6. Data abstraction issuc for cdgc computing Fig. 5. Naming mechanism in edgeS used for things management in edges. Network address such as IP address or MAC address will be used to support various well. However, they are not flexible enough to serve the communication protocols such as BlueTooth, Zig Bee or WiFi dynamic edge network since sometimes most of the things and so on. When targeting highly dynamic environment such t edge could be highly mobile and resource constrained as city level system, we think it is still an open problem and Moreover, for some resource constrained things at the edge worth further investigation by the community of the network, ip based naming scheme could be too heavy to support considering its complexity and overhead New naming mechanisms such as named data network- C. Data Abstraction ing(NDN)[27] and Mobility First [28] could also be applied Various applications can run on the edgeS consuming to edge computing. ndn provide a hierarchically structured data or providing service by communicating through the air name for content/data centric network, and it is human friendly position indicators from the service management layer. Data for service management and provides good scalability for abstraction has been well discussed and researched in the wire- edge. However, it would need extra proxy in order to fit into less sensor network and cloud computing paradigm. However, other communication protocols such as BlueTooth or Zig Bee, in edge computing, this issue becomes more challenging. With and so on. Another issue associated with ndn is security, IoT, there would be a huge number of data generators in the since it is very hard to isolate things hardware information network, and here we take a smart home environment as an with service providers. MobileFirst can separate name from example. In a smart home, almost all of the things will report network address in order to provide better mobility support, data to the edgeS, not to mention the large number of things and it would be very efficient if applied to edge services where deployed all around the home. However, most of the things at things are of highly mobility. Neverless, a global unique iden- the edge of the network, only periodically report sensed data tification (GUID)needs to be used for naming is MobileFirst, to the gateway. For example, the thermometer could report the and this is not required in related fixed information aggregation temperature every minute, but this data will most likely only service at the edge of the network such as home environment. be consumed by the real user several times a day. another Another disadvantage of Mobile First for edge is the difficulty example could be a security camera in the home which might in service management since guid is not human friendly keep recording and sending the video to the gateway, but the For a relative small and fixed edge such as home environ- data will just be stored in the database for a certain time with ment, let the edgeOS assign network address to each thing nobody actually consuming it, and then be flushed by the latest could be a solution with in one system, each thing could have video a unique human friendly name which describes the following Based on this observation, we envision that human involve information: location(where), role (who), and data descrip- ment in edge computing should be minimized and the edge tion(what), for example, kitchen oven temperature. Then node should consume/process all the data and interact with the edges will assign identifier and network address to this users in a proactive fashion. In this case, data should be prepro thing, as shown in Fig. 5. The human friendly name is unique cessed at the gateway level, such as noise/low-quality removal for each thing and it will be used for service management, event detection, and privacy protection, and so on. Processed things diagnosis, and component replacement. For user and data will be sent to the upper layer for future service providing service provider, this naming mechanism makes management There will be several challenges in this process very easy. For example, the user will receive a message from First, data reported from different things comes with var edgeS like"Bulb 3(what)of the ceiling light (who) in living ious formats, as shown in Fig. 6. For the concern of room(where) failed, and then the user can directly replace privacy and security. applications running on the gateway the failed bulb without searching for an error code or recon- should be blinded from raw data. Moreover, they should figure the network address for the new bulb. Moreover, this extract the knowledge they are interested in from an inte- naming mechanism provides better programmability to service grated data table. We can easily define the table with providers and in the meanwhile, it blocks service providers id, time, name, data (e. g, 10000, 12: 34: 56PM 01701/2016 from getting hardware information, which will protect data pri- kitchen. oven2. temperature, 78|)such that any edge things vacy and security better. Unique identifier and network address data can be fitted in. However, the details of sensed data have could be mapped from human friendly name Identifier will be been hidden, which may affect the usability of data SHI et al.: EDGE COMPUTING: VISION AND CHALLENGES Second, it is sometimes difficult to decide the degree of detected by the Os before an application is installed, then data abstraction If too much raw data is filtered out, some a user can be warned and avoid the potential access issue applications or services could not learn enough knowledge. Another side of the isolation challenge is how to isolate a However, if we want to keep a large quantity of raw data, there user's private data from third party applications. For exam would be a challenge for data storage. Lastly, data reported ple, your activity tracking application should not be able to y things at edge could be not reliable sometime, due to access your electricity usage data. To solve this challenge, a the low precision sensor, hazard environment, and unreliable well-designed control access mechanism should be added to wireless connection. In this case, how to abstract useful infor- the service management layer in the edges mation from unreliable data source is still a challenge for Iot Reliabilily: Last but not least, reliability is also a key chal- application and system developers lenge at the edge of the network. We identify the challenges One more issue with data abstraction is the applicable opera- in reliability from the different views of service, system, and tions on the things. Collecting data is to serve the application data here and the application should be allowed to control (e. g, read 1) From the service point of view, it is sometimes very hard from and write to) the things in order to complete certain ser to identify the reason for a service failure accurately at vices the user desires. Combining the data representation and field oIe, if an air conditioner is not working operations, the data abstraction layer will serve as an public potential reason could be that a power cord is cut, com- interface for all things connected to edges. Furthermore, due pressor failure or even a temperature controller has run the heterogeneity of the things both data representation and out of battery. A sensor node could have lost connec allowed operations could diverse a lot, which also increases tion very easily to the system due to battery outage, bad the barrier of universal data abstraction connection condition, component wear out, etc. At the edge of the network, it is not enough to just maintain a current service when some nodes lose connection but to D. Service management provide the action after node failure makes more sense In terms of service management at the edge of the net to the user. For example, it would be very nice if the work, we argue that the following four fundamental features edges could inform the user which component in the should be supported to guarantee a reliable system, including service is not responding, or even alert the user ahead differentiation, extensibility, isolation, and reliability if some parts in the system have a high risk of failure Differentiation: With the fast growth of lot deployment, Potential solutions for this challenge could be adapted we expected multiple services will be deployed at the edge from a wireless sensor network. or industrial network of the network such as smart home. These services will such as PROFINET [29 have different priorities. For example, critical services such as 2) From the system point of view, it is very important for things diagnosis and failure alarm should be processed earlier the edges to maintain the network topology of the than ordinary service Health related service, for example, fall whole system, and each component in the system is detection or heart failure detection should also have a higher able to send status/diagnosis information to the edgeS priority compared with other service such as entertainment With this feature, Services such as failure detection, thing Extensibility: Extensibility could be a huge challenge at the replacement, and data quality detection could be easily edge of the network, unlike a mobile system, the things in the deployed at the system level loT could be very dynamic. When the owner purchases a new 3) From the data point of view, reliability challenge rise thing, can it be easily added to the current service without any mostly from the data sensing and communication part problem? Or when one thing is replaced due to wearing out, As previously researched and discussed, things at the can the previous service adopt a new node easily? These prob edge of the network could fail due to various reasons and lems should be solved with a flexible and extensible design of they could also report low fidelity data under unreliable service management layer in the edges condition such as low battery level [30]. Also various Isolation: Isolation would be another issue at the edge of new communication protocols for lot data collection the network. In mobile OS, if an application fails or crashes are also proposed. These protocols serves well for the the whole system will usually crash and reboot. Or in a dis support of huge number of sensor nodes and the highly tributed system the shared resource could be managed with dynamic network condition [31. However, the connec different synchronization mechanisms such as a lock or token tion reliability is not as good as blueTooth or WiF ring. However, in a smart edgeS, this issue might be more If both sensing data and communication are not reli complicated. There could be several applications that share able, how the system can still provide reliable service by the same data resource, for example. the control of light. If leveraging multiple reference data source and historical one application failed or was not responding, a user should data record is still an open challenge still be able to control their lights without crashing the whole edgeOS. Or when a user removes the only application that controls lights from the system, the lights should still be alive E. Privacy and securily rather than experiencing a lost connection to the edgeOS. At the edge of the network, usage privacy and data secu This challenge could be potentially solved y protection are the most important services that should b deployment/undeployment framework. If the conflict could be provided. If a home is deployed with loT,a lot of privacy IEEE INTERNET OF THINGS JOURNAL. VOL 3. NO. 5 OCTOBER 2016 information can be learned from the sensed usage data. For Latency: Latency is one of the most important metrics to example, with the reading of the electricity or water usage, evaluate the performance, especially in interaction applica- one can easily speculate if the house is vacant or not. In this tions/services [341,[35]. Servers in cloud computing provide case, how to support service without harming privacy is a high computation capability. They can handle complex work- challenge. Some of the private information could be removed loads in a relatively short time, such as image processing, from data before processing such as masking all the faces in voice recognition and so on. However, latency is not only the video. We think that keeping the computing at the edge determined by computation time. Long Wan delays can of data resource, which means in the home, could be a decent dramatically influence the real-time/interaction intensive app method to protect privacy and data security. To protect the data cations behavior [36]. To reduce the latency, the workload security and usage privacy at the edge of the network, several should better be finished in the nearest layer which has enough challenges remain open computation capability to the things at the edge of the network First is the awareness of privacy and security to the commu- For example, in the smart city case, we can leverage phones nity. We take WiFi networks security as an example. Among to process their local photos first then send a potential missing the 439 million households who use wireless connections, child's into back to the cloud instead of uploading all photos 49%0 of WiFi networks are unsecured, and 80% of house- Due to the large amount of photos and their size, it will be holds still have their routers set on default passwords. For much faster to preprocess at the edge. However, the nearest public WiFi hotspots, 89%0 of them are unsecured [32]. All physical layer may not always be a good option. We need to the stake holders including service provider, system and appli- consider the resource usage information to avoid unnecessar cation developer and end user need to aware that the users' waiting time so that a logical optimal layer can be found. If privacy would be harmed without notice at the edge of the net- user is playing games, since the phones computation resource work. For example, ip camera, health monitor, or even some is already occupied, it will be better to upload a photo to the WiFi enabled toys could easily be connected by others if not nearest gateway or micro-center protected properl Bandwidth: From latency's point of view, high bandwidth Second is the ownership of the data collected from things at can reduce transmission time, especially for large data(e. g edge. Just as what happened with mobile applications, the data video, etc. )[37, [38]. For short distance transmission, we can of end user collected by things will be stored and analyzed at establish high bandwidth wireless access to send data to the the service provider side. However, leave the data at the edge edge. On one hand, if the workload can be handled at the where it is collected and let the user fully own the data will be edge, the latency can be greatly improved compared to work a better solution for privacy protection. Similar to the health on the cloud. The bandwidth between the edge and the cloud record data, end user data collected at the edge of the network is also saved For example, in the smart home case, almost all should be stored at the edge and the user should be able to the data can be handled in the home gateway through Wi-Fi or control if the data should be used by service providers. During other high speed transmission methods. In addition, the trans- the process of authorization, highly private data could also be mission reliability is also enhanced as the transmission path is removed by the things to further protect user privacy short. On the other hand. although the transmission distance Third is the missing of efficient tools to protect data pri- cannot be reduced since the edge cannot satisfy the computa vacy and security at the edge of the network. Some of the tion demand, at least the data is preprocessed at the edge and things are highly resource constrained so the current meth- the upload data size will be significantly reduced. In the smart ods for security protection might not be able to be deployed city case, it is better to preprocess photos before upload, so on thing because they are resource hungry. moreover, the the data size can be greatly reduced. It saves the users' band highly dynamic environment at the edge of the network also width, especially if they are using a carriers data plan. From makes the network become vulnerable or unprotected. For a global perspective, the bandwidth is saved in both situa- privacy protection, some platform such as Open mHealth is tions, and it can be used by other edges to upload/download proposed to standardize and store health data [33 but more data. Hence, we need to evaluate if a high bandwidth con- tools are still missing to handle diverse data attributes for edge nection is needed and which speed is suitable for an edge computin Besides, to correctly determine the workload allocation in each layer, we need to consider the computation capability and F. Optimization Metrics handwidth usage information in layers to avoid competition In edge computing, we have multiple layers with different and delay computation capability. Workload allocation becomes a big Energy: Battery is the most precious resource for things issue. We need to decide which layer to handle the workload at the edge of the network. For the endpoint layer, offload or how many tasks to assign at each part. There are multi- ing workload to the edge can be treated as an energy free ple allocation strategies to complete a workload, for instances, method [22,[39]. So for a given workload, is it energy effi evenly distribute the workload on each layer or complete as cient to offload the whole workload (or part of it) to the edge much as possible on each layer. The extreme cases are fully rather than compute locally? The key is the tradeoff between operated on endpoint or fully operated on cloud. To choose the computation energy consumption and transmission energy an optimal allocation strategy, we discuss several optimization consumption. Generally speaking we first need to consider the metrics in this section, including latency, bandwidth, energy power characteristics of the workload. Is it computation inten- and cost sive How much resource will it use to run locally besides the SHI et al.: EDGE COMPUTING: VISION AND CHALLENGES 645 network signal strength [401, the data size and available band- conventional cloud computing paradigm still supported, but width will also influence the transmission energy overhead also it can connect long distance networks together for data [28]. We prefer to use edge computing only if the transmis- sharing and collaboration because of the closeness of data. At sion overhead is smaller than computing locally. However, if last, we put forward the challenges and opportunities that are we care about the whole edge computing process rather than worth working on, including programmability, naming, data only focus on endpoints, total energy consumption should be abstraction, service management, privacy and security, as well the accumulation of each used layers energy cost. Similar to as optimization metrics. edge computing is here, and we hope the endpoint layer, each layer's energy consumption can be this paper will bring this to the attention of the community estimated as local computation cost plus transmission cost. In this case the optimal workload allocation strategy may change ACKNOWLEdgment For example, the local data center layer is busy, so the work- load is continuously uploaded to the upper layer. Comparing The authors would like to thank T. Zhang from Cisco for with computing on endpoints, the multihop transmission may early discussions and w. Zhang from Alibaba for the idea of dramatically increase the overhead which causes more energy edge computing and fog computing. The example of apply consumption ing a shopping cart at the edge was given by w. Zhang. The Cost: From the service providers perspective, e. g authors would also like to thank Dr C. Wang for inviting them You Tube, Amazon, etc, edge computing provides them less to submit this paper latency and energy consumption, potential increased through put and improved user experience. As a result, they can earn REFERENCES more money for handling the same unit of workload. For [1] M. Armbrust et al., "A view of cloud computing "Commun. ACM example, based on most residents'interest, we can put a pop vol.53、no.4,pp.50-58,2010 ular video on the building layer edge. The city layer edge can 2] S. Ghemawat, H. Gobioff, and S.-T.Leung, "The Google file system, ACM SIGOPS Oper Syst. Rev.. vol 37, no 5, pp 29-43, 2003 free from this task and handle more complex work. The total [3] J. Dean and s. Ghemawat, "MapReduce: Simplified data processing throughput can be increased. The investment of the service on large clusters, Commun. ACM, vol. 51, no. 1, pp. 107-113, 2008 providers is the cost to build and maintain the things in each [4]K. Shvachko, H. Kuang, s Radia, and R Chansler, " The hadoop dis layer. To fully utilize the local data in each layer, providers tributed file system,in Proc. IEEE 26th Symp. Mass Storage Syst Technol(MSST), Inclinc Village, NV, USA, 2010, pp. 1-10 can charge users based on the data location. New cost models [5] M. Zaharia, M. Chowdhury, M. J. Franklin, S. Shenker, and I. Stoica, need to be developed to guarantee the profit of the service Spark: Cluster computing with working sets, in Proc. 2nd USENIX provider as well as acceptability of users Con/. Hot Topics Cloud Comput., vol. 10. Bostoll, MA, USA, 2010, P.10 Workload allocation is not an easy task. The metrics are [6] K. Ashton, "That Internet of Things thing, "RFID J, vol. 22, no. 7 closely related to each other. For example, due to the energy pp.97-114,2009 constraints, a workload needs to be complete on the city data [7 H. Sunldimlaeker, P. Guillemin, P. Friess, and s. Woelme, ""Vision and challenges for realising the Internet of things, voL. 20, no. 10, 2010 center layer. Comparing with the building server layer, the 8 J. Gubbi, R. Buyya, s. Marusic, and M. Palaniswami, "Internet of energy limitation inevitably affects the latency. Metrics should Things (oT): A vision, architectural elements, and future directions be given priority (or weight) for different workloads so that Future Gener: Comput. Syst., vol. 29, no. 7, pp. 1645-1660, 2013 19]"Cisco global cloud index: Forecast and methodology, 2014-2019 a reasonable allocation strategy can be selected. Besides, the white paper, 2014 cost analysis needs to be done in runtime. The interference and [Io] D. Evans, The Internet of Things: How the next evolution of the resource usage of concurrent workloads should be considered Internet is changing everything, "CISCO White Paper, vol. 1, pp. 1-11 as well lIl D. Evans, The Internet of Things: How the next evolution of the Internet is changing everything, CISCO White Paper, vol. 1, pp. 1-11 V. CONCLUSION [12]A. Greenberg, J. Hamilton, D. A Maltz, and P. Patel, "The cost of a cloud: Research problems in data center networks, ACM SIGCOMM Nowadays, more and more services are pushed from the Comput. Commun. Rev., vol 39, no. I, pp 68-73, 2008 cloud to the edge of the network because processing data at 13 E. Cuervo et al.,"MAUl: Making smartphones last longer with code offload, in Proc. &th Int. Conf. MObile Syst. Appl. Services the edge can ensure shorter response time and better reliability San Francisco, CA, USA, 2010, pp. 49-62 Moreover, bandwidth could also be saved if a larger portion [14] M. Satyanarayanan, P Bahl, R Caceres, and N Davies, "The case for of data could be handled at the edge rather than uploaded to VM-based cloudlets in mobile computing, IEEE Pervasive Comput vol 8 no. 4. pp. 14-23. Oct /ec. 2009 the cloud. The burgeoning of loT and the universalized mobile [15] F. Bonomi, R. Milito, J. Zhu, and S. Addepalli, "Fog computing and devices changed the role of edge in the computing paradigm its role in the Internet of things, in Proc. Ist Edition MCC Workshop from data consumer to data producer/consumer. It would b Mobile Cloud Comput., Helsinki, Finland, 2012, pp. 13-16 [16] Boeing 787s to Create Half a Terabyte of Data Per more efficient to process or massage data at the edge of the Flight, Says Virgin Atlantic. Accessed on Dec. 7, 2016 network. In this paper, we came up with our understanding Online].Available:https://datafloq.com/read/self-driving-cars- of edge computing, with the rationale that computing should create-2-petabytes-data-annually/172 happen at the proximity of data sources. Then we list [17 Self-Driving Cars Will Create 2 Petabytes of Data. What are the Big Data Opportunities for the Ca eral cases whereby edge computing could flourish from cloud Industry? Accessed on Dec. 7, 2016. [Online offloading to a smart environment such as home and city. We Availahle:http://www.computerworlduk.com/news/data/boeing-787 also introduce collaborative edge, since edge can connect end create-half-terabyte-of-data-per-flight-says-virgin-atlantic-3433595. [18 Data Never Sleeps 2. 0. Accessed on Dec 7, 2016 [ Online]. Available user and cloud both physically and logically so not only is the https://www.domo.com/blog/2014/04/data-never-sleeps-2-0/ 646 IEEE INTERNET OF THINGS JOURNAL. VOL 3. NO. 5 OCTOBER 2016 [19](2016). Open Fog Architecture Overview. Open Fog Consortium Weisong Shi rcccivcd thc Ph.D. dcgrcc in computcr Architecture Working Group. Accessed on Dec. 7, 2016 engineering from the Chinese Academy of Sciences, [oNline].Available:http://www.openfogconsortium.org/wp-content/ Beijing, China. uploads/OpenFog-Architecture-Overview-WP-2-2016 pdf He is a Professor of computer science with Wayne Technol(Hor Web), Washington, DC, USA, 2015, pp. 73-0 veb JFo [ 20]S. Yi, Z. Hao, Z. Qin, and Q. Li, "Fog computing: Platform and State University, Detroit, MI, USA. His current applications, in Proc. 3rd IEEE Workshop Hot Topics research interests include big data systems, ed computing, energy-efficient computer systems, and 21]K. Ila et al., "Towards wearable cognitive assistance. in Proc. 12th software and mobile health Annu. Int. Conf. Mobile Syst. Appl. Services, Bretton Woods, NH, USA, Dr. Shi is a senior member of the acm 14.pp.68-81. [22] B.-G. Chun, S Ihm, P Maniatis, M. Naik, and A. Patti, Clone Cloud Elastic execution between mobile device and cloud, in Proc. 6th Conf Comput. Syst, Salzburg, Austria, 2011, pp 301-314 [23]A. Rudenko, P. Reiher, G. J. Popek, and G. H. Kuenning, Saving portable computer battery power through remote process execution, ACM SIGMOBILE Mobile Comput. Commun. Rev., vol. 2, no. I pp.19-26,1998 Jie Cao(S16)received the B.S. degree in telecom munications engineering from Xidian University. [241 G. C. Hunt and M. L. Scott, The coign automatic distributed parti tioning systcm, in Proc. OSDI, vol 99 Ncw Orlcans, LA, USA, 1999. Xian, China. He is currently pursuing the Ph.D pp.187-200 degree in computer science at Wayne State University. Detroit Ml usa [25] K. Kumar and Y.H. Lu. "Cloud computing for mobile users: Can offloading computation save energy? "'Computer, voL. 43, n0. 4. His current research interests include edge com- p.5156,Apr.2010. puting, smart homc, mobilc, and connccted health [26]s. Kosta, A. Aucinas, P. Hui, R. Mortier, and X. /hang, ThinkAir: Dynamic resource allocation and parallel execution in the cloud for mobilc codc offloading, in Proc. IEEE INFOCOM, Orlando, FL, USA, [27] L Zhang et al., "Named data networking (NDN) project, Xerox Palo Alto Res Center, Palo Allo, CA, USA, Tech. Rep. NDN-0001, 2010 [28] D Raychaudhuri, K. Nagaraja, and A. Venkataramani, ""Mobility First A robust and trustworthy mobility-centric architecture for the future Internet, ACM SIGMOBILE MObile Comput. Commun. Rev., voL 16 Quan Zhang (S16 received the B.s. degree in computer science from Tongji University, Shanghai 229]J. Feld, " PROFINET-Scalable factory communication for all applica- China. He is currently pursuing the Ph. D. degree in computer science at Wayne State University, Detroit, Austria, 2004, Pp 33-38 ML USA [30]J. Cao, L. Ren, w Shi, and Z. Yu, "A framework for component selec His current research interests include distributed tion in collaborative sensing application development. " in Proc. 10th systcms, cloud computing, and cncrgy-cfficicnt com- IEEE Conf. Coll. Comput. Netw. Appl. Worksharing, Miami, FL. USA puting 2014,Pp.104-113 [31] F. DaCosta, Rethinking the Internet of Things: A Scalable Approach to C Everything. New York, NY, USA: A 2013 [321 WiFi Network Security Statistics/Graph. Accessed on Dec. 7, 2016 [oNlinc].Availablc:http:/graphs.nct/wifi-stats.html/ 33] Open Mhealth Platform. Accessed on Dec. 7, 2016. [Online Available:http:/www.openmhealth.org/ [34]K.R. Jackson et al., ""Performance analysis of high performance com- puting applications on the Amazon Web services cloud, in Proc. IEEE Youhuizi Li received the: Ph. D. degree in computer 2nd Int. Conf. Cloud Comput. Technol. Sci. (Cloud Com), Indianapoli science from Wayne State University, Detroit, MI IN,USA,2010.pp.159168 USA. in 2016 [35]A. Li, X. Yang, S. Kandula, and M. Zhang,"CloudCmp She is an Assistant Professor of computer sc Comparing public cloud providers, in Proc. 10th ACM SIGCOMM ence with Hangzhou Dianzi University. Hangzhou Conf. Internet Meas, 2010, pp 1-14. IOnline]. Available China. Her current research interests include energy http://doi.acmorg/10.1145/1879141.1879143 cfficicnt computing, mobilc and Intcrnct computing [36] M. Satyanarayanan, " Mobile computing: The next decade and big data systems SIGMOBILE Mobile Comput. Commun. Rev. no.2.pp.2-10,2011.[oNline].Available:http://doi.acmorg/ 10.1145/2016598.2016600 37 A. Greenberg, J. Hamilton, D. A. Maltz, and P. Patel, " The cost of a cloud: Research problems in data center networks, SIGCOMM Comput. Commun. Rev, vol 39, no. 1, pp 68-73, Dec 2008. Online Available:http://doi.acm.org/10.1145/1496091.1496103 [38] M. Armbrust et al.,A view of cloud computing? Commun ACM, voL. 53, no. 4. pp. 50-58, Apr. 2010. [Online]. Available Lanyu Xu received the b.s. degree in computer sci http://doi.acm.org/10.1145/1721654.1721672 ence from Tongji University, Shanghai, China. She is [39A. P. Miettinen and J. K. Nurminen. Energy efficiency of mobile currently pursuing the Ph. D. degree in computer sci- clients in cloud computing, in Proc. 2nd (SENIX Conf. Hot Topics ence at Wayne State University, Detroit, MI, USA Cloud Compul, Boston, MA, USA, 2010, p. 4.Online Available Her current research interests include edge com- http://dl.acm.org/citation.cfm?id=1863103.1863107 puting, smart homes, and mobile and connect health [40] N. Ding et al., " Characterizing and modeling the impact of wireless signal strength on smartphone battery drain, SIGMETRICS Perform Eval. Rev, vol. 41, nU. l, pp. 29-40, Jull. 2013. [Online]. Available http://doi.acmorg/10.1145/2494232.2466586

(系统自动生成,下载前可以参看下载内容)