COVER FEATURE OUTLOOKThe Emergenceof Edge ComputingMahadev Satyanarayanan, Carnegie Mellon UniversityIndustry investment and research interest in edge computing, inwhich computing and storage nodes are placed at the Internet’sedge in close proximity to mobile devices or sensors, have growndramatically in recent years. This emerging technology promisesto deliver highly responsive cloud services for mobile computing,scalability and privacy-policy enforcement for the Internet ofThings, and the ability to mask transient cloud outages.Cloud computing, which has dominated ITdiscourse in the past decade, has a twofoldvalue proposition. First, centralization exploitseconomies of scale to lower the marginalcost of system administration and operations. Second,organizations can avoid the capital expenditure ofcreating a datacenter by consuming computing resourcesover the Internet from a large service provider. Theseconsiderations have led to the consolidation of computingcapacity into multiple large datacenters spread across theglobe. The proven economic benefits of cloud computingmake it likely to remain a permanent feature of the futurecomputing landscape.However, the forces driving centralization are not theonly ones at work. Nascent technologies and applicationsfor mobile computing and the Internet of Things (IoT) aredriving computing toward dispersion. Edge computing is a30CO M PUTE R P U B LISHED BY THE IEEE COMP UTER SOCIE T Y new paradigm in which substantial computing and storageresources—variously referred to as cloudlets,1 microdatacenters, or fog nodes2—are placed at the Internet’sedge in close proximity to mobile devices or sensors.Industry investment and research interest in edgecomputing have grown dramatically in recent years.Nokia and IBM jointly introduced the Radio ApplicationsCloud Server (RACS), an edge computing platform for4G/LTE networks, in early 2013.3 The following year, amobile edge computing standardization effort beganunder the auspices of the European TelecommunicationsStandards Institute (ETSI).4 The Open Edge Computinginitiative (OEC; openedgecomputing.org) was launched inJune 2015 by Vodafone, Intel, and Huawei in partnershipwith Carnegie Mellon University (CMU) and expandeda year later to include Verizon, Deutsche Telekom,T-Mobile, Nokia, and Crown Castle. This collaboration0 0 1 8 - 9 1 6 2 / 1 7/ 3 3 .0 0 2 0 1 7 I E E E
includes creation of a Living EdgeLab in Pittsburgh, Pennsylvania, togain hands-on experience with a livedeployment of proof-of-concept cloudletbased applications. Organized by thetelecom industry, the first Mobile EdgeComputing Congress (tmt.knect365.com/mobile-edge-computing) convened inLondon in September 2015 and againin Munich a year later. The Open FogConsortium (www.openfogconsortium.org) was created by Cisco, Microsoft,Intel, Dell, and ARM in partnershipwith Princeton University in November2015, and has since expanded to includemany other companies. The First IEEE/ACM Symposium on Edge Computing(conferences.computer.org/SEC) was heldin October 2016 in Washington, DC.These developments raise severalquestions: why has edge computingemerged, what new capabilities does itenable, and where is it headed?ORIGIN AND BACKGROUNDThe roots of edge computing reachback to the late 1990s, when Akamaiintroduced content delivery networks(CDNs) to accelerate web performance.5A CDN uses nodes at the edge close tousers to prefetch and cache web content.These edge nodes can also perform somecontent customization, such as addinglocation-relevant advertising. CDNs areespecially valuable for video content,because the bandwidth savings fromcaching can be substantial.Edge computing generalizes andextends the CDN concept by leveragingcloud computing infrastructure. Aswith CDNs, the proximity of cloudletsto end users is crucial. However,instead of being limited to caching webcontent, a cloudlet can run arbitrarycode just as in cloud computing. Thiscode is typically encapsulated in avirtual machine (VM) or a lighter-weightcontainer for isolation, safety, resourcemanagement, and metering.In 1997, Brian Noble and hiscolleagues first demonstrated edgecomputing’s potential value to mobilecomputing.6 They showed how speechClearly, reliance on a cloud datacenteris not advisable for applications thatrequire end-to-end delays to be tightlycontrolled to less than a few tens ofmilliseconds. As will be discussed later,tight control of latency is necessaryUSING PERSISTENT CACHING SIMPLIFIESTHE MANAGEMENT OF CLOUDLETSDESPITE THEIR PHYSICAL DISPERSAL ATTHE INTERNET EDGE.recognition could be implementedwith acceptable performance on aresource-limited mobile device byoffloading computation to a nearbyserver. Two years later, Jason Flinnand I extended this approach toimprove battery life.7 In a 2001 articlethat generalized these concepts, Iintroduced the term cyber foraging forthe amplification of a mobile device’scomputing capabilities by leveragingnearby infrastructure.8Cloud computing’s emergence in themid-2000s led to the cloud becoming themost obvious infrastructure to leveragefrom a mobile device. Today, Apple’sSiri and Google’s speech-recognitionservices both offload computation tothe cloud. Unfortunately, consolidationimplies large average separationbetween a mobile device and its optimalcloud datacenter. Ang Li and hiscolleagues reported that the averageround-trip time from 260 globalvantage points to their optimal AmazonElastic Compute Cloud (EC2) instancesis 74 ms.9 To this must be added thelatency of a wireless first hop. In termsof jitter, the variance inherent in amultihop network must be included.for emerging applications such asaugmented reality (AR).These observations about end-toend latency and cloud computingwere first articulated in a 2009 articleI coauthored with Paramvir Bahl,Rámon Cáceres, and Nigel Daviesthat laid the conceptual foundationfor edge computing.1 We advocateda two-level architecture: the firstlevel is today’s unmodified cloudinfrastructure; the second levelconsists of dispersed elements calledcloudlets with state cached from thefirst level. Using persistent cachinginstead of hard state simplifies themanagement of cloudlets despite theirphysical dispersal at the Internet edge.The cloudlet concept can, of course,be expanded to a multilevel cloudlethierarchy.In 2012, Flavio Bonomi and hiscolleagues introduced the term fogcomputing to refer to this dispersedcloud infrastructure.2 However, theirmotivation for decentralization isIoT infrastructure scalability ratherthan mobile applications’ interactiveperformance. The researchers envisiona multilevel hierarchy of fog nodesJANUARY 2017 31
OUTLOOK1.00.80.60.40.203.3 J1.1 J3.1 J5.1 J5.2 J9.4 J0200(a)400600800Mobile onlyCloudletAWS-EastAWS-WestAWS-EUAWS-Asia16.4 J5.4 J6.6 J8.5 J9.5 J14.3 econds8001,000(b)FIGURE 1. Response time distribution and per-operation energy cost of an (a) augmented reality and (b) face recognition application ona 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 adaptedfrom 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 edgedevices.WHY PROXIMITY MATTERSAs we explore new applications anduse cases for both mobile computingand the IoT, the virtues of proximityare becoming increasingly apparent.In the physical world, the importanceof proximity has never been indoubt. The old axiom about thethree top determinants of real estatevalue being “location, location, andlocation” captures this observationwell. In the cyber world, the seamlessconnectivity offered by the Internethas lulled us into a false sense ofdisregard for physical proximity.Because logical network proximity isentirely characterized by low latency,low jitter, and high bandwidth, thequestion “How close is physicallyclose enough?” cannot be answered inthe abstract. It is dependent on factorssuch as the networking technologiesused, network contention, applicationcharacteristics, and user tolerance forpoor interactive response.Physical proximity affects end-to- end latency, economically viableband w idth, establishment of trust,and survivability. With sufficienteffort and resource investment, thelack of proximity can be partiallymasked. For example, a direct fiberconnection can achieve low latencyand high bandwidth between distantpoints. However, there are limits tothis approach. The speed of light is anobvious physical limit on latency. The32COMPUTER need to use a multihop networkingstrategy to cover a large geographicarea with many access points imposesan economic limit on both latencyand bandwidth. Each hop introducesqueuing and routing delay, as well asbuffer bloat.10The proximity of cloudlets helps inat least four distinct ways:›› Highly responsive cloud services.A cloudlet’s physical proximity to a mobile device makes iteasier to achieve low end-to-endlatency, high bandwidth, andlow jitter to services located onthe cloudlet. This is valuable forapplications such as AR and virtual reality that offload computation to the cloudlet.›› Scalability via edge analytics. Thecumulative ingress bandwidthdemand into the cloud from alarge collection of high-bandwidth IoT sensors, such as videocameras, is considerably lowerif the raw data is analyzedon cloudlets. Only the (muchsmaller) extracted informationand metadata must be transmitted to the cloud.›› Privacy-policy enforcement. Byserving as the first point ofcontact in the infrastructure forIoT sensor data, a cloudlet canenforce the privacy policies of itsowner prior to release of the datato the cloud.›› Masking cloud outages. If a cloudservice becomes unavailable dueto network failure, cloud failure,or a denial-of-service attack,a fallback service on a nearbycloudlet can temporarily maskthe failure.I now discuss each of these advantages in detail.HIGHLY RESPONSIVECLOUD SERVICESHumans are acutely sensitive to delaysin the critical path of interaction, andtheir performance on cognitive tasksis remarkably fast and accurate.11For example, under normal lightingconditions, face recognition takes370–620 ms, depending on familiarity.Speech recognition takes 300–450 msfor short phrases, and it requires only4 ms to tell that a sound is a humanvoice. VR applications that use headtracked systems require latencies ofless than 16 ms to achieve perceptualstability. End-to-end latency of a fewtens of milliseconds is a safe butachievable goal.Figure 1 illustrates the importanceof cloudlets for low-latency offloadservices. The graphs show thecumulative distribution of measuredresponse times for an AR and a facerecognition application on a mobiledevice.12 An image from the mobiledevice, which is located in Pittsburgh,is transmitted over a Wi-Fi firsthop to a cloudlet or to an AmazonWeb Services (AWS) datacenter. Theimage is processed at the destinationby computer vision code executingW W W.CO M P U T E R .O R G /CO M P U T E R
within a VM. For AR, buildings inthe image are recognized and labelscorresponding to their identities aretransmitted back to the mobile device.For face recognition, the identity of theperson is returned.The ideal curve in Figure 1 wouldbe a step function that jumps to 1.0at the origin. As the figure shows,the ideal is best approximated by acloudlet. End-to-end network latencyimpedes performance, as indicated bythe worsening response-time curvescorresponding to more distant AWSlocations. Increasing response timealso increases per-operation energyconsumption on the mobile device.This value is indicated beside thecorresponding label in the figurelegend. For example, the deviceconsumes 1.1 J on average to performan AR operation on the cloudlet, but3.1 J, 5.1 J, and so on when performingit on AWS-East, AWS-West, and so on.Similar results can be expected withany offload service that is concentratedin a few large datacenters.The label “mobile only” in thefigure corresponds to a case whereno offloading is performed and thecomputer vision code is run on themobile device. In spite of avoidingthe energy and performance cost ofWi-Fi communication, this optionis slower than using the cloudlet.Offloading is clearly important forthese applications.Cloudlets are a disruptive technologythat brings energy-rich high-end computing within one wireless hop ofmobile devices, thereby enabling newapplications that are both computationintensive and latency-sensitive. A primeexample is wearable cognitive assistance,11which combines a device like GoogleGlass with cloudlet-based processingto guide users through a complex task.As with a GPS system, the user hearsa synthesized voice describing whatto do next and sees visual cues in theGlass display. The system catches errorsimmediately and corrects the userbefore they cascade. The final report ofthe 2013 National Science Foundationsmaller task-specific state space. Thesecond phase of each task workflowoperates solely on the symbolic representation. Comparing the symbolicrepresentation to the expected taskstate generates user guidance for thenext step (last column of Table 1). TheINDEPENDENT OF LATENCYCONSIDERATIONS, CLOUDLETSCAN REDUCE INGRESS BANDWIDTHINTO THE CLOUD.Workshop on Future Directions inWireless Networking characterizedthis new genre of applications as“astonishingly transformative.”13 Inongoing work at CMU,14 we have builtcognitive assistance applications forthe seven tasks summarized in Table 1.Videos of some of these applications areavailable at goo.gl/02m0nL.On the cloudlet, the workflow ofthese applications consists of twophases. In the first phase, the sensorinputs are analyzed to extract a symbolic representation of task progress(fourth column of Table 1). This is anidealized representation of the inputsensor values relative to the task,and excludes all irrelevant detail.This phase must be tolerant of considerable real-world variability—forexample, different lighting levels,light sources, viewer’s positions withrespect to the task artifacts, task- unrelated clutter in the background,and so on. One can view the extractionof a symbolic representation as atask-specific “analog-to-digital” conversion: the enormous state space ofsensor values is simplified to a muchvideo guidance is shown on the Glassdisplay, and audio guidance is givenusing the Android text-to-speech API.SCALABILITY THROUGHEDGE ANALYTICSIndependent of latency considerations,cloudlets can also reduce ingressbandwidth into the cloud. For example,consider an application in which manycolocated users are continuouslytransmitting video from their smartphone to the cloud for contentanalysis. The cumulative data ratefor even a small fraction of users ina modest-size city would saturate itsmetropolitan area network: 12,000users transmitting 1080p video wouldrequire a link of 100 gigabits persecond; a million users would require alink of 8.5 terabits per second.Figure 2 shows how cloudlets cansolve this problem. In the proposedGigaSight framework,15 video froma mobile device only travels as far asa nearby cloudlet. The cloudlet runscomputer vision analytics in nearreal time and only sends the results(content tags, recognized faces, andJANUARY 2017 33
OUTLOOKTABLE 1. Example wearable cognitive assistance applications.AppnameExample input videoframeApp descriptionSymbolicrepresentationExample guidanceFaceJogs user’s memory of a familiar face whose name cannotbe recalled. Detects and extracts a tightly cropped imageof each face, then applies popular open source facerecognizer OpenFace (cmusatyalab.github.io/openface),which is based on a deep neural network (DNN) algorithm.Whispers name of person. Can be used in combination withmood detection algorithms to offer conversational hints.ASCII text of nameWhispers “BarackObama”PoolHelps novice pool player aim correctly. Gives continuousvisual feedback (left arrow, right arrow, or thumbs up) asuser turns cue stick. Correct shot angle is calculated basedon widely used fractional aiming system. Uses color, line,contour, and shape detection. Symbolic representationdescribes positions of cue ball, object ball, target pocket,and top and bottom of cue stick. Pocket, object ball,cue ball, cue top, cuebottom PingPongTells novice player to hit ball to left or right, depending onwhich is more likely to beat opponent. Uses color, line, andoptical-flow-based motion detection to detect ball, table,and opponent. Symbolic representation is a 3-tuple: in rallyor not, opponent position, ball position. Whispers “left” or“right.” InRally, ball position,opponent position WorkoutGuides correct user form in exercise actions like sit-ups Action count and push-ups, and counts out repetitions. Uses volumetrictemplate matching on a 10- to 15-frame video segment toclassify poorly performed repetitions as distinct types ofexercise (for example, “bad push-up”). Uses smartphone onfloor for third-person viewpoint.LegoGuides user in assembling 2D Lego models. Analyzeseach video frame in three steps: (1) finds board using itsdistinctive color and black dot pattern, (2) locates Legobricks on board using edge and color detection, and(3) assigns brick color using weighted majority votingwithin each block. Symbolic representation is matrixshowing color for each brick.DrawHelps user to sketch better. Builds on third-party apporiginally designed to input sketches from pen tabletsand output corrective guidance on desktop screen. Ourimplementation preserves back-end logic. New GoogleGlass–based front end allows use of any drawing surfaceand instrument and displays guidance on Glass. Displayserror alignment in sketch.SandwichHelps cooking novice prepare sandwiches according to arecipe. Because real food is perishable, we use realisticplastic toy food as ingredients. Object detection uses aregion proposal and DNN approach. Implementation is ontop of Caffe (caffe.berkeleyvision.org) and Dlib (dlib.net).Transfer learning saves time in labeling and training.Whispers “Left”Says “8”[[ 0, 2, 1, 1 ],[ 0, 2, 1, 6 ],[ 2, 2, 2, 2 ]]Says “Find a 1 3green piece and put iton top”Object: “Lettuce on topof ham and bread”Says “Now put apiece of bread on thelettuce”34COMPUTERW W W.CO M P U T E R .O R G /CO M P U T E R
BlankingSamplingUser 1EncoderPrivacysettingsEarly discardVideoDecoderMetadataContent filterPersonal VM 1BlurringUser 2Reduce up-front work sample video frames only denature samples rest of video is encrypted per-VM encryption key decrypt on demand only index nt filterVideoDecoderMetadataEarly discardPersonal VM 2BlurringIndexer VMTagging ofdenatured videoCloudlet storageFIGURE 2. GigaSight framework. A cloudlet performs computer vision analytics on video from mobile devices in near real time and onlysends the results along with metadata to the cloud, sharply reducing ingress bandwidth into the cloud. VM: Virtual machine.so on) along with metadata (owner,capture location, timestamp, and so on)to the cloud. This can reduce ingressbandwidth into the cloud by three tosix orders of magnitude. GigaSightalso shows how tags and metadata inthe cloud can guide deeper and morecustomized searches of the contentof a video segment during its (finite)retention period on a cloudlet.A video camera is only one exampleof a high-data-rate sensor in theIoT. Another example is a modernaircraft, which can generate nearlyhalf a terabyte of sensor data duringa flight. Real-time analysis of thisdata on a cloudlet in the aircraft couldgenerate timely guidance for preventivemaintenance, fuel economy, and otherbenefits.Cloudlets’ latency and bandwidthadvantages are especially relevantin the context of automobiles, tocomplement vehicle-to-vehicle (V2V)approaches bein
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
May 02, 2018 · D. Program Evaluation ͟The organization has provided a description of the framework for how each program will be evaluated. The framework should include all the elements below: ͟The evaluation methods are cost-effective for the organization ͟Quantitative and qualitative data is being collected (at Basics tier, data collection must have begun)
Silat is a combative art of self-defense and survival rooted from Matay archipelago. It was traced at thé early of Langkasuka Kingdom (2nd century CE) till thé reign of Melaka (Malaysia) Sultanate era (13th century). Silat has now evolved to become part of social culture and tradition with thé appearance of a fine physical and spiritual .
On an exceptional basis, Member States may request UNESCO to provide thé candidates with access to thé platform so they can complète thé form by themselves. Thèse requests must be addressed to esd rize unesco. or by 15 A ril 2021 UNESCO will provide thé nomineewith accessto thé platform via their émail address.
̶The leading indicator of employee engagement is based on the quality of the relationship between employee and supervisor Empower your managers! ̶Help them understand the impact on the organization ̶Share important changes, plan options, tasks, and deadlines ̶Provide key messages and talking points ̶Prepare them to answer employee questions
Dr. Sunita Bharatwal** Dr. Pawan Garga*** Abstract Customer satisfaction is derived from thè functionalities and values, a product or Service can provide. The current study aims to segregate thè dimensions of ordine Service quality and gather insights on its impact on web shopping. The trends of purchases have
Chính Văn.- Còn đức Thế tôn thì tuệ giác cực kỳ trong sạch 8: hiện hành bất nhị 9, đạt đến vô tướng 10, đứng vào chỗ đứng của các đức Thế tôn 11, thể hiện tính bình đẳng của các Ngài, đến chỗ không còn chướng ngại 12, giáo pháp không thể khuynh đảo, tâm thức không bị cản trở, cái được
Le genou de Lucy. Odile Jacob. 1999. Coppens Y. Pré-textes. L’homme préhistorique en morceaux. Eds Odile Jacob. 2011. Costentin J., Delaveau P. Café, thé, chocolat, les bons effets sur le cerveau et pour le corps. Editions Odile Jacob. 2010. Crawford M., Marsh D. The driving force : food in human evolution and the future.
Le genou de Lucy. Odile Jacob. 1999. Coppens Y. Pré-textes. L’homme préhistorique en morceaux. Eds Odile Jacob. 2011. Costentin J., Delaveau P. Café, thé, chocolat, les bons effets sur le cerveau et pour le corps. Editions Odile Jacob. 2010. 3 Crawford M., Marsh D. The driving force : food in human evolution and the future.