COVER FEATURE OUTLOOK
The Emergence of Edge Computing
Mahadev Satyanarayanan, Carnegie Mellon University
Industry investment and research interest in edge computing, in which computing and storage nodes are placed at the Internet’s edge in close proximity to mobile devices or sensors, have grown dramatically in recent years. This emerging technology promises to deliver highly responsive cloud services for mobile computing, scalability and privacy-policy enforcement for the Internet of Things, and the ability to mask transient cloud outages.
loud computing, which has dominated IT new paradigm in which substantial computing and storage discourse in the past decade, has a twofold resources—variously referred to as cloudlets,1 micro value proposition. First, centralization exploits datacenters, or fog nodes2—are placed at the Internet’s economies of scale to lower the marginal edge in close proximity to mobile devices or sensors. Ccost of system administration and operations. Second, Industry investment and research interest in edge organizations can avoid the capital expenditure of computing have grown dramatically in recent years. creating a datacenter by consuming computing resources Nokia and IBM jointly introduced the Radio Applications over the Internet from a large service provider. These Cloud Server (RACS), an edge computing platform for considerations have led to the consolidation of computing 4G/LTE networks, in early 2013.3 The following year, a capacity into multiple large datacenters spread across the mobile edge computing standardization effort began globe. The proven economic benefits of cloud computing under the auspices of the European Telecommunications make it likely to remain a permanent feature of the future Standards Institute (ETSI).4 The Open Edge Computing computing landscape. initiative (OEC; openedgecomputing.org) was launched in However, the forces driving centralization are not the June 2015 by Vodafone, Intel, and Huawei in partnership only ones at work. Nascent technologies and applications with Carnegie Mellon University (CMU) and expanded for mobile computing and the Internet of Things (IoT) are a year later to include Verizon, Deutsche Telekom, driving computing toward dispersion. Edge computing is a T-Mobile, Nokia, and Crown Castle. This collaboration
30 COMPUTER PUBLISHED BY THE IEEE COMPUTER SOCIETY 0018-9162/17/$33.00 © 2017 IEEE includes creation of a Living Edge container for isolation, safety, resource Clearly, reliance on a cloud datacenter Lab in Pittsburgh, Pennsylvania, to management, and metering. is not advisable for applications that gain hands-on experience with a live In 1997, Brian Noble and his require end-to-end delays to be tightly deployment of proof-of-concept cloudlet- colleagues first demonstrated edge controlled to less than a few tens of based applications. Organized by the computing’s potential value to mobile milliseconds. As will be discussed later, telecom industry, the first Mobile Edge computing.6 They showed how speech tight control of latency is necessary Computing Congress (tmt.knect365.com /mobile-edge-computing) convened in London in September 2015 and again in Munich a year later. The Open Fog USING PERSISTENT CACHING SIMPLIFIES Consortium (www.openfogconsortium .org) was created by Cisco, Microsoft, THE MANAGEMENT OF CLOUDLETS Intel, Dell, and ARM in partnership DESPITE THEIR PHYSICAL DISPERSAL AT with Princeton University in November THE INTERNET EDGE. 2015, and has since expanded to include many other companies. The First IEEE/ ACM Symposium on Edge Computing (conferences.computer.org/SEC) was held recognition could be implemented for emerging applications such as in October 2016 in Washington, DC. with acceptable performance on a augmented reality (AR). These developments raise several resource-limited mobile device by These observations about end-to- questions: why has edge computing offloading computation to a nearby end latency and cloud computing emerged, what new capabilities does it server. Two years later, Jason Flinn were first articulated in a 2009 article enable, and where is it headed? and I extended this approach to I coauthored with Paramvir Bahl, improve battery life.7 In a 2001 article Rámon Cáceres, and Nigel Davies ORIGIN AND BACKGROUND that generalized these concepts, I that laid the conceptual foundation The roots of edge computing reach introduced the term cyber foraging for for edge computing.1 We advocated back to the late 1990s, when Akamai the amplification of a mobile device’s a two-level architecture: the first introduced content delivery networks computing capabilities by leveraging level is today’s unmodified cloud (CDNs) to accelerate web performance.5 nearby infrastructure.8 infrastructure; the second level A CDN uses nodes at the edge close to Cloud computing’s emergence in the consists of dispersed elements called users to prefetch and cache web content. mid-2000s led to the cloud becoming the cloudlets with state cached from the These edge nodes can also perform some most obvious infrastructure to leverage first level. Using persistent caching content customization, such as adding from a mobile device. Today, Apple’s instead of hard state simplifies the location-relevant advertising. CDNs are Siri and Google’s speech-recognition management of cloudlets despite their especially valuable for video content, services both offload computation to physical dispersal at the Internet edge. because the bandwidth savings from the cloud. Unfortunately, consolidation The cloudlet concept can, of course, caching can be substantial. implies large average separation be expanded to a multilevel cloudlet Edge computing generalizes and between a mobile device and its optimal hierarchy. extends the CDN concept by leveraging cloud datacenter. Ang Li and his In 2012, Flavio Bonomi and his cloud computing infrastructure. As colleagues reported that the average colleagues introduced the term fog with CDNs, the proximity of cloudlets round-trip time from 260 global computing to refer to this dispersed to end users is crucial. However, vantage points to their optimal Amazon cloud infrastructure.2 However, their instead of being limited to caching web Elastic Compute Cloud (EC2) instances motivation for decentralization is content, a cloudlet can run arbitrary is 74 ms.9 To this must be added the IoT infrastructure scalability rather code just as in cloud computing. This latency of a wireless first hop. In terms than mobile applications’ interactive code is typically encapsulated in a of jitter, the variance inherent in a performance. The researchers envision virtual machine (VM) or a lighter-weight multihop network must be included. a multilevel hierarchy of fog nodes
JANUARY 2017 31 OUTLOOK
1.0 1.0 0.8 3.3 J Mobile only 16.4 J 0.8 1.1 J Cloudlet 5.4 J 0.6 3.1 J AWS-East 6.6 J 0.6 0.4 5.1 J AWS-West 8.5 J 0.4 0.2 5.2 J AWS-EU 9.5 J 0.2 0 9.4 J AWS-Asia 14.3 J 0 0 200 400 600 800 1,000 0 200 400 600 800 1,000 Milliseconds Milliseconds (a) (b)
FIGURE 1. Response time distribution and per-operation energy cost of an (a) augmented reality and (b) face recognition application on a mobile device, in which an image from the device is transmitted over a Wi-Fi first hop to a cloudlet or an Amazon Web Services (AWS) datacenter. The ideal is best approximated by a cloudlet, demonstrating the importance of low-latency offload services. Figure adapted from K. Ha et al., “The Impact of Mobile Multimedia Applications on Data Center Consolidation,” Proc. 2013 IEEE Int’l Conf. Cloud Eng. (IC2E 13), 2013, pp. 166–176.
stretching from the cloud to IoT edge need to use a multihop networking to network failure, cloud failure, devices. strategy to cover a large geographic or a denial-of-service attack, area with many access points imposes a fallback service on a nearby WHY PROXIMITY MATTERS an economic limit on both latency cloudlet can temporarily mask As we explore new applications and and bandwidth. Each hop introduces the failure. use cases for both mobile computing queuing and routing delay, as well as and the IoT, the virtues of proximity buffer bloat.10 I now discuss each of these advan- are becoming increasingly apparent. The proximity of cloudlets helps in tages in detail. In the physical world, the importance at least four distinct ways: of proximity has never been in HIGHLY RESPONSIVE doubt. The old axiom about the ››Highly responsive cloud services. CLOUD SERVICES three top determinants of real estate A cloudlet’s physical proxim- Humans are acutely sensitive to delays value being “location, location, and ity to a mobile device makes it in the critical path of interaction, and location” captures this observation easier to achieve low end-to-end their performance on cognitive tasks well. In the cyber world, the seamless latency, high bandwidth, and is remarkably fast and accurate.11 connectivity offered by the Internet low jitter to services located on For example, under normal lighting has lulled us into a false sense of the cloudlet. This is valuable for conditions, face recognition takes disregard for physical proximity. applications such as AR and vir- 370–620 ms, depending on familiarity. Because logical network proximity is tual reality that offload compu- Speech recognition takes 300–450 ms entirely characterized by low latency, tation to the cloudlet. for short phrases, and it requires only low jitter, and high bandwidth, the ››Scalability via edge analytics. The 4 ms to tell that a sound is a human question “How close is physically cumulative ingress bandwidth voice. VR applications that use head- close enough?” cannot be answered in demand into the cloud from a tracked systems require latencies of the abstract. It is dependent on factors large collection of high-band- less than 16 ms to achieve perceptual such as the networking technologies width IoT sensors, such as video stability. End-to-end latency of a few used, network contention, application cameras, is considerably lower tens of milliseconds is a safe but characteristics, and user tolerance for if the raw data is analyzed achievable goal. poor interactive response. on cloudlets. Only the (much Figure 1 illustrates the importance Physical proximity affects end-to- smaller) extracted information of cloudlets for low-latency offload end latency, economically viable and metadata must be transmit- services. The graphs show the bandwidth, establishment of trust, ted to the cloud. cumulative distribution of measured and survivability. With sufficient ››Privacy-policy enforcement. By response times for an AR and a face effort and resource investment, the serving as the first point of recognition application on a mobile lack of proximity can be partially contact in the infrastructure for device.12 An image from the mobile masked. For example, a direct fiber IoT sensor data, a cloudlet can device, which is located in Pittsburgh, connection can achieve low latency enforce the privacy policies of its is transmitted over a Wi-Fi first and high bandwidth between distant owner prior to release of the data hop to a cloudlet or to an Amazon points. However, there are limits to to the cloud. Web Services (AWS) datacenter. The this approach. The speed of light is an ››Masking cloud outages. If a cloud image is processed at the destination obvious physical limit on latency. The service becomes unavailable due by computer vision code executing
32 COMPUTER WWW.COMPUTER.ORG/COMPUTER within a VM. For AR, buildings in As with a GPS system, the user hears smaller task-specific state space. The the image are recognized and labels a synthesized voice describing what second phase of each task workflow corresponding to their identities are to do next and sees visual cues in the operates solely on the symbolic rep- transmitted back to the mobile device. Glass display. The system catches errors resentation. Comparing the symbolic For face recognition, the identity of the immediately and corrects the user representation to the expected task person is returned. before they cascade. The final report of state generates user guidance for the The ideal curve in Figure 1 would the 2013 National Science Foundation next step (last column of Table 1). The be a step function that jumps to 1.0 at the origin. As the figure shows, the ideal is best approximated by a cloudlet. End-to-end network latency INDEPENDENT OF LATENCY impedes performance, as indicated by the worsening response-time curves CONSIDERATIONS, CLOUDLETS corresponding to more distant AWS CAN REDUCE INGRESS BANDWIDTH locations. Increasing response time INTO THE CLOUD. also increases per-operation energy consumption on the mobile device. This value is indicated beside the corresponding label in the figure Workshop on Future Directions in video guidance is shown on the Glass legend. For example, the device Wireless Networking characterized display, and audio guidance is given consumes 1.1 J on average to perform this new genre of applications as using the Android text-to-speech API. an AR operation on the cloudlet, but “astonishingly transformative.”13 In 3.1 J, 5.1 J, and so on when performing ongoing work at CMU,14 we have built SCALABILITY THROUGH it on AWS-East, AWS-West, and so on. cognitive assistance applications for EDGE ANALYTICS Similar results can be expected with the seven tasks summarized in Table 1. Independent of latency considerations, any offload service that is concentrated Videos of some of these applications are cloudlets can also reduce ingress in a few large datacenters. available at goo.gl/02m0nL. bandwidth into the cloud. For example, The label “mobile only” in the On the cloudlet, the workflow of consider an application in which many figure corresponds to a case where these applications consists of two colocated users are continuously no offloading is performed and the phases. In the first phase, the sensor transmitting video from their smart- computer vision code is run on the inputs are analyzed to extract a sym- phone to the cloud for content mobile device. In spite of avoiding bolic representation of task progress analysis. The cumulative data rate the energy and performance cost of (fourth column of Table 1). This is an for even a small fraction of users in Wi-Fi communication, this option idealized representation of the input a modest-size city would saturate its is slower than using the cloudlet. sensor values relative to the task, metropolitan area network: 12,000 Offloading is clearly important for and excludes all irrelevant detail. users transmitting 1080p video would these applications. This phase must be tolerant of con- require a link of 100 gigabits per Cloudlets are a disruptive technology siderable real-world variability—for second; a million users would require a that brings energy-rich high-end com- example, different lighting levels, link of 8.5 terabits per second. puting within one wireless hop of light sources, viewer’s positions with Figure 2 shows how cloudlets can mobile devices, thereby enabling new respect to the task artifacts, task- solve this problem. In the proposed applications that are both computation- unrelated clutter in the background, GigaSight framework,15 video from intensive and latency-sensitive. A prime and so on. One can view the extraction a mobile device only travels as far as example is wearable cognitive assistance,11 of a symbolic representation as a a nearby cloudlet. The cloudlet runs which combines a device like Google task-specific “analog-to-digital” con- computer vision analytics in near Glass with cloudlet-based processing version: the enormous state space of real time and only sends the results to guide users through a complex task. sensor values is simplified to a much (content tags, recognized faces, and
JANUARY 2017 33 OUTLOOK
TABLE 1. Example wearable cognitive assistance applications.
App Example input video App description Symbolic Example guidance name frame representation
Face Jogs user’s memory of a familiar face whose name cannot ASCII text of name Whispers “Barack be recalled. Detects and extracts a tightly cropped image Obama” of each face, then applies popular open source face recognizer OpenFace (cmusatyalab.github.io/openface), which is based on a deep neural network (DNN) algorithm. Whispers name of person. Can be used in combination with mood detection algorithms to o er conversational hints.
Pool Helps novice pool player aim correctly. Gives continuous
Ping- Tells novice player to hit ball to left or right, depending on
Workout Guides correct user form in exercise actions like sit-ups
Lego Guides user in assembling 2D Lego models. Analyzes [[ 0, 2, 1, 1 ], each video frame in three steps: (1) fi nds board using its [ 0, 2, 1, 6 ], distinctive color and black dot pattern, (2) locates Lego [ 2, 2, 2, 2 ]] bricks on board using edge and color detection, and (3) assigns brick color using weighted majority voting within each block. Symbolic representation is matrix Says “Find a 1 × 3 showing color for each brick. green piece and put it on top”
Draw Helps user to sketch better. Builds on third-party app originally designed to input sketches from pen tablets and output corrective guidance on desktop screen. Our implementation preserves back-end logic. New Google Glass–based front end allows use of any drawing surface and instrument and displays guidance on Glass. Displays error alignment in sketch.
Sandwich Helps cooking novice prepare sandwiches according to a Object: “Lettuce on top recipe. Because real food is perishable, we use realistic of ham and bread” plastic toy food as ingredients. Object detection uses a region proposal and DNN approach. Implementation is on top of Ca e (ca e.berkeleyvision.org) and Dlib (dlib.net). Transfer learning saves time in labeling and training.
Says “Now put a piece of bread on the lettuce”
34 COMPUTER WWW.COMPUTER.ORG/COMPUTER Personal VM 1
Metadata Video Encoder Privacy Decoder Early discard Content lter settings Sampling Blanking User 1 Blurring
Personal VM 2
Metadata Video
Privacy Encoder Decoder Early discard settings Content lter User 2 Sampling Blanking Blurring
Reduce up-front work Indexer VM • sample video frames • only denature samples • rest of video is encrypted • per-VM encryption key • decrypt on demand • only index samples Tagging of denatured video Cloudlet storage
FIGURE 2. GigaSight framework. A cloudlet performs computer vision analytics on video from mobile devices in near real time and only sends the results along with metadata to the cloud, sharply reducing ingress bandwidth into the cloud. VM: Virtual machine.
so on) along with metadata (owner, approaches being explored for real-time might reveal model-specific defects capture location, timestamp, and so on) control and accident avoidance. For the that could be corrected in a timely to the cloud. This can reduce ingress foreseeable future, cloud connectivity manner. bandwidth into the cloud by three to from a moving automobile will at For other automotive applications, six orders of magnitude. GigaSight best be 3G or 4G/LTE. An important such as collaborative real-time avoidance also shows how tags and metadata in question is whether cloudlets should of road hazards, the telco cloudlet is the cloud can guide deeper and more be in automobiles or part of the telco the optimal hosting site. For example, customized searches of the content infrastructure (perhaps one cloudlet if a vehicle hits a pothole or swerves to of a video segment during its (finite) connected via fiber links to multiple avoid a fallen tree branch, the hazard’s retention period on a cloudlet. cell towers in an area). Both alternatives coordinates can be rapidly shared A video camera is only one example have value. within the telco cloudlet and then used of a high-data-rate sensor in the An application such as a multiplayer for many hours by other automobiles IoT. Another example is a modern video game for automobile passengers to proactively cope with the hazard aircraft, which can generate nearly is best hosted on the vehicle’s cloudlet. (for example, by warning drivers to half a terabyte of sensor data during The cloudlet could also perform make an early lane change). a flight. Real-time analysis of this real-time analytics of high-data-rate data on a cloudlet in the aircraft could sensor streams from the engine and PRIVACY POLICY generate timely guidance for preventive other sources to alert the driver to ENFORCEMENT maintenance, fuel economy, and other imminent failure or the need for Cloudlets could help address a benefits. preventive maintenance. In addition, vexing problem—namely, growing Cloudlets’ latency and bandwidth this information could be transmitted concerns over data privacy arising advantages are especially relevant to the cloud for integration into the from IoT system overcentralization. in the context of automobiles, to vehicle manufacturer’s database. Fine- Increasingly reluctant to release raw complement vehicle-to-vehicle (V2V) grain analysis of such anomaly data sensor data to an IoT cloud hub, users
JANUARY 2017 35 OUTLOOK
IoT hub in the cloud
Denatured Full- delity sensor request streams
Cloudlet
Video cameras
Privacy Privacy mediator 2 mediator 3 All raw sensor for sensors for sensors data (for example, from home Privacy P, Q, R X, Y, Z or local organization) mediator 1 for sensors A, B, C, D
Motion sensors Thermal sensors
Linux kernel + Cloudlet manager
Local cloudlet storage ( nite history of full- delity sensor data)
Electric power Water- ow sensors sensors
FIGURE 3. Internet of Things (IoT) privacy architecture. Software modules called privacy mediators execute on a cloudlet within a sensor owner’s trust domain to perform denaturing and privacy-policy enforcement on the sensor streams.
and organizations desire finer-grain infrastructure contact for the sensor denatured data. Whether to relax control over release of that data. For streams. Trusted software modules the normal privacy policy in such example, a user should be able to called privacy mediators execute on situations is a decision that remains delete or denature a subset of sensor the cloudlet to perform denaturing under end-user control. data he or she deems sensitive. From and privacy-policy enforcement the end user’s per-spective, denatured on the sensor streams. Cloudlets MASKING CLOUD OUTAGES sensor data is safe to release to the thus provide the foundation for a As our dependence on the cloud grows, outside world: faces in images can scalable and secure privacy solution so does our vulnerability to cloud be blurred, sensor readings can be that aligns well with natural outages. Implicit in the convergence coarsely aggregated or omitted at organizational boundaries of trust of mobile and cloud computing is the certain times of day or night, and so and responsibility. assumption that the cloud is easily on. Today’s IoT architectures, in which As Figure 3 illustrates, full-fidelity accessible at all times—in other words, data is transmitted directly from sensor data can be archived on a there is good end-to-end network quality sensors to a cloud hub, make such fine- cloudlet for a finite duration such as a and few network or cloud failures. grain control impossible. few hours, days, or weeks depending However, there are usage contexts in Nigel Davies and his colleagues16 on data volume and local storage which cloud access must be viewed as propose an IoT privacy architecture size. This could be valuable in case an occasional luxury rather than a basic (see Figure 3) that leverages a cloudlet the IoT hub discovers an anomaly and necessity. This viewpoint applies to within a sensor owner’s trust domain. returns a request for more in-depth several important contexts that can be This cloudlet is the first point of data analysis using less aggressively referred to as hostile environments.
36 COMPUTER WWW.COMPUTER.ORG/COMPUTER The prime example of a hostile updates to the cloud, and detecting and On the nontechnical side, the environment is a theater of military resolving conflicts). Generalizing these biggest unknown relates to viable operations—jamming the enemy’s net- steps to various cloud services will be business models for deploying work is a standard tactic. Another an important future research area. cloudlets. Success will require the example is a geographical region where recovery is under way after networking infrastructure has been destroyed by a natural disaster or terrorist attack. A third example DURING FAILURES, A CLOUDLET CAN is a developing country with weak SERVE AS A PROXY FOR THE CLOUD AND networking infrastructure. A fourth PERFORM ITS CRITICAL SERVICES. example is any part of the Internet that has temporarily become a hostile environment because it is under cyber- attack. There is growing concern that cyberattacks could soon become major THE ROAD AHEAD involvement and support of a complex weapons of organized crime as well as Edge computing clearly offers many set of industries, communities, and instruments of national policy. If these benefits. At the same time, it also faces standards organizations. This presents dire predictions come true, the entire many technical and nontechnical a classic bootstrapping problem. Internet might have to be viewed as a challenges. Without unique applications and hostile environment in the future. On the technical side, there are services that leverage edge computing, Cloudlets can alleviate cloud out- many unknowns pertaining to the there is no incentive for deploying ages. Because of physical proximity, software mechanisms and algorithms cloudlets. Yet, without large-enough the survivability characteristics of a needed for the collective control and cloudlet deployments, there is little cloudlet are closer to its associated sharing of cloudlets in distributed incentive for developers to create mobile devices than to the distant computing. There are also substantial those new applications and services. cloud. This opens the door to hurdles in managing dispersed cloudlet How can we break this deadlock? approaches in which a fallback service infrastructure. As mentioned earlier, This state of affairs is similar to on a cloudlet can temporarily mask one of cloud computing’s driving that at the dawn of the Internet in the cloud inaccessibility.17 During failures, forces is the lower management cost late 1970s to early 1980s. An open a cloudlet can serve as a proxy for the of centralized infrastructure. The ecosystem attracted investment in cloud and perform its critical services. dispersion inherent in edge computing infrastructure and applications, with- Upon repair of the failure, actions that raises the complexity of management out any single entity bearing large were tentatively committed to the considerably. Developing innovative risk or dominating the market. Over cloudlet might need to be propagated to technical solutions to reduce this time, this lead to the emergence of a the cloud for reconciliation. complexity is a research priority for critical mass of Internet infrastructure More than two decades ago, James edge computing. Another important and applications (such as email) that Kistler and I anticipated how this area of study will be the development could uniquely benefit from that concept could be applied to cloud- of mechanisms to compensate for the infrastructure. By the time the World sourced data in describing the Coda File weaker perimeter security of cloudlets, Wide Web emerged as a “killer app” System, which provided disconnectable relative to cloud datacenters. The in the early 1990s, sufficient Internet read–write access to shared data.18 The development of tamper-resistant and infrastructure had been deployed for essential steps are hoarding (prefetching tamper-evident enclosures, remote growth to explode. data into a persistent cache), emulation surveillance, and Trusted Platform Edge computing can follow a (leveraging hoarded data in the cloud’s Module–based attestation are all similar, but faster, path to success absence and precisely tracking local important paths that could contribute by nurturing the creation of an open updates), and reintegration (propagating to alleviating this problem. cloudlet ecosystem. This is the goal of
JANUARY 2017 37 OUTLOOK
mid-2000s, the centralized approach of cloud computing began its ascent to ABOUT THE AUTHOR the preeminent position that it holds today. Edge computing represents the MAHADEV SATYANARAYANAN (Satya) is the Carnegie Group Professor latest phase of this ongoing dialectic. of Computer Science at Carnegie Mellon University (CMU). His multidecade research career has focused on the challenges of performance, scalability, ACKNOWLEDGMENTS availability, and trust in information systems that reach from the cloud to the The ideas and results presented in this mobile edge of the Internet. In the course of this work, he pioneered many article have arisen from my discussions advances in distributed systems, mobile computing, pervasive computing, the and research collaborations with Internet of Things, and, most recently, edge/fog computing. Satya received a many individuals over the past decade: PhD in computer science from CMU. He is a Fellow of ACM and IEEE. Contact Yoshihisa Abe, Brandon Amos, Victor him at [email protected]. Bahl, Vas Bala, Jeff Boleng, Ramón Cáceres, Zhuo Chen, Sarah Clinch, Nigel Davies, Khalid Elgazzar, Roxana Geambasu, Benjamin Gilbert, Adam Goode, Kiryong Ha, Jan Harkes, Martial Hebert, Wenlu OEC’s OpenStack++, a derivative of the of 5G networks because it ensures Hu, Canturk Isci, Kaustubh Joshi, Guenter popular OpenStack cloud computing that ultra-low first-hop latency is not Klas, Bobby Klatzky, Grace Lewis, Ed platform. The “++” refers to the unique swamped by the much larger latency Morris, Padmanabhan Pillai, Wolfgang extensions necessary for cloudlet of the remaining hops to the cloud. A Richter, Bill Schilit, Rolf Schuster, Dan environments including cloudlet third trend is continuing improvement Siewiorek, Soumya Simanta, Pieter discovery, just-in-time provisioning, in the computing capabilities of Simoens, Nina Taft, Brandon Taylor, and VM hand-off. As edge computing wearables, smartphones, and other Junjue Wang, Yu Xiao, and Roy Want. grows, OpenStack++ aims to become mobile devices that represent the I also would like to thank the following a widely used platform that catalyzes Internet’s extreme edge. Although people who provided valuable feedback on many proprietary and nonproprietary these devices are indeed growing in an early draft of the article and helped innovations in hardware, software, computing power, their improvements to improve it: Zhuo Chen, Guenter Klas, and services. are muted by the fundamental Padmanabhan Pillai, Rolf Schuster, The emergence of edge computing challenges of mobility such as weight, Weisong Shi, Marco Silva, Stu Wagner, coincides with three important trends size, battery life, and heat dissipation. and the anonymous reviewers. in the computing and communication The sweet spot for edge computing is This work was supported by the landscape that, despite being driven thus in the infrastructure, where it can National Science Foundation under grant by distinct forces, are convergent. One amplify the capabilities of proximate numbers IIS-1065336 and CNS-1518865. trend is software-defined networking mobile devices and sensors. Additional support was provided by (SDN) and the associated concept Intel, Vodafone, Google, Crown Castle, of network function virtualization and the Conklin Kistler family fund. (NFV), which must be supported n closing, it is useful to reflect on Any opinions, findings, conclusions, by some of the same virtualized edge computing from a historical or recommendations expressed in this infrastructure as edge computing. A perspective. Since the 1960s, com- material are mine and should not be second trend is growing interest in Iputing has alternated between attributed to Carnegie Mellon University ultra-low-latency (one millisecond centralization and decentralization. or the funding sources. or less) wireless networks for a new The centralized approaches of batch class of tactile applications. Ultra- processing and timesharing prevailed REFERENCES low latency is one of the proposed in the 1960s and 1970s. The 1980s and 1. M. Satyanarayanan et al., “The attributes for future 5G networks. 1990s saw decentralization through Case for VM-Based Cloudlets in Edge computing is a natural partner the rise of personal computing. By the Mobile Computing,” IEEE Pervasive
38 COMPUTER WWW.COMPUTER.ORG/COMPUTER Computing, vol. , no. , , . J. Flinn and M. Satyanarayanan, Cognitive Assistance Applications,” pp. – . “Energy-Aware Adaptation for Proc. Workshop Wearable Sys- . F. Bonomi et al., “Fog Computing and Mobile Applications,” Proc. th ACM tems and Applications (WearSys ), Its Role in the Internet of Things,” Symp. Operating Systems Principles , pp. – . Proc. st Edition MCC Workshop (SOSP ), , pp. –. . P. Simoens et al., “Scalable Mobile Cloud Computing (MCC ), . M. Satyanarayanan, “Pervasive Com- Crowd-Sourcing of Video from , pp. –. puting: Vision and Challenges,” IEEE Mobile Devices,” Proc. th Int’l Conf. . “IBM and Nokia Siemens Networks Personal Comm., vol. , no. , , Mobile Systems, Applications, and Ser- Announce World’s First Mobile Edge pp. –. vices (MobiSys ), , pp. – . Computing Platform MWC,” press . A. Li et al., “CloudCmp: Comparing . N. Davies et al., “Privacy Mediators: release, Nokia, Feb. ; company Public Cloud Providers,” Proc. th Helping IoT Cross the Chasm,” Proc. .nokia.com/en/news/press-releases ACM SIGCOMM Conf. Internet Mea- th Int’l Workshop Mobile Computing / / / /ibm-and-nokia-siemens surement (IMC ), , pp. – . Systems and Applications (HotMobile -networks-announce-worlds- rst . J. Gettys and K. Nichols, “Bu erbloat: ), , pp. – . -mobile-edge-computing-platform Dark Bu ers in the Internet,” ACM . M. Satyanarayanan et al., “The Role -mwc. Queue, vol. , no. , ; queue.acm of Cloudlets in Hostile Environ- . S. Antipolis, “ETSI Announces First .org/detail.cfm?id= . ments,” IEEE Pervasive Computing, Meeting of New Standardization . K. Ha et al., “Towards Wearable vol. , no. , , pp. – . Group on Mobile-Edge Computing,” Cognitive Assistance,” Proc. th Int’l . J.J. Kistler and M. Satyanarayanan, The Standard, European Telecommu- Conf. Mobile Systems, Applications, and “Disconnected Operation in the Coda nications Standards Inst., Oct. Services (MobiSys ), , pp. – . File System,” ACM Trans. Computer ; www.etsi.org/news-events . K. Ha et al., “The Impact of Mobile Systems, vol. , no. , , pp. – . /news/ - - -news-etsi Multimedia Applications on Data -announces- rst-meeting-of-new Center Consolidation,” Proc.
-standardization-group-on-mobile IEEE Int’l Conf. Cloud Eng. (IC E ), See www.computer.org -edge-computing. , pp. –. /computer-multimedia . J. Dilley et al., “Globally Distributed . S. Banerjee and D.O. Wu, Final Report for multimedia content related to this article. Content Delivery,” IEEE Internet Com- from the NSF Workshop on Future puting, vol. , no. , , pp. – . Directions in Wireless Networking, . B.D. Noble et al., “Agile Application- Nat’l Science Foundation, Nov. ; Aware Adaptation for Mobility,” ecedha.org/docs/nsf-nets/ nal Read your subscriptions Proc. th ACM Symp. Operating -report.pdf. through the myCS publications portal at Systems Principles (SOSP ), , . Z. Chen et al., “Early Implemen- http://mycs.computer.org pp. – . tation Experience with Wearable
Subscribe today for the latest in computational science and engineering research, news and analysis, CSE in education, and emerging technologies in the hard sciences. www.computer.org/cise
JANUARY 2017 39