VIRTUAL MACHINES The Case for VM-Based Cloudlets in Mobile Computing A new vision of mobile computing liberates mobile devices from severe resource constraints by enabling resource-intensive applications to leverage cloud computing free of WAN delays, jitter, congestion, and failures. obile computing is at a to this transformation and proposes a new ar- fork in the road. After two chitecture for overcoming them. In this archi- decades of sustained effort tecture, a mobile user exploits virtual machine by many researchers, we’ve (VM) technology to rapidly instantiate custom- finally developed the core ized service software on a nearby cloudlet and Mconcepts, techniques, and mechanisms to pro- then uses that service over a wireless LAN; the vide a solid foundation for this still fast-growing mobile device typically functions as a thin cli- area. The vision of “information at my finger- ent with respect to the service. A cloudlet is a tips at any time and place” was just a dream in trusted, resource-rich computer or cluster of the mid 1990s; today, ubiquitous email and Web computers that’s well-connected to the Internet access is a reality that millions of users world- and available for use by nearby mobile devices. wide experience through BlackBerries, iPhones, Our strategy of leveraging transiently cus- Windows Mobile, and other mobile devices. On tomized proximate infrastructure as a mobile one path of the fork, mobile Web-based services device moves with its user through the physical and location-aware advertising opportunities world is called cloudlet-based, resource-rich, have begun to appear, and companies are mak- mobile computing. Crisp interactive response, ing large investments in antici- which is essential for seamless augmentation Mahadev Satyanarayanan pation of major profits. of human cognition, is easily achieved in this Carnegie Mellon University Yet, this path also leads mo- architecture because of the cloudlet’s physical bile computing away from its proximity and one-hop network latency. Using Paramvir Bahl true potential. Awaiting dis- a cloudlet also simplifies the challenge of meet- Microsoft Research covery on the other path is an ing the peak bandwidth demand of multiple us- Ramón Cáceres entirely new world in which ers interactively generating and receiving media AT&T Research mobile computing seamlessly such as high-definition video and high-resolu- augments users’ cognitive tion images. Rapid customization of infrastruc- Nigel Davies abilities via compute-intensive ture for diverse applications emerges as a critical Lancaster University capabilities such as speech requirement, and our results from a proof-of- recognition, natural language concept prototype suggest that VM technology processing, computer vision can indeed help meet this requirement. and graphics, machine learning, augmented re- ality, planning, and decision-making. By thus Resource-Poor Mobile Hardware empowering mobile users, we could transform The phrase “resource-rich mobile comput- many areas of human activity (see the sidebar ing” seems like an oxymoron at first glance. for an example). Researchers have long recognized that mobile This article discusses the technical obstacles hardware is necessarily resource-poor relative 2 PERVASIVE computing Published by the IEEE CS n 1536-1268/09/$26.00 © 2009 IEEE Help for the Mentally Challenged magine a future in which there are extensive deployments Ron with cognitive assistance. At the heart of this technology I of dense cloudlet infrastructure based on open standards, is a lightweight wearable computer with a head-up display much like Wi-Fi access points today. What kind of new appli- in the form of eyeglasses. Built into the eyeglass frame are a cations can we envision in such a world? camera for scene capture and earphones for audio feedback. These hardware components offer the essentials of an aug- Ron has recently been diagnosed with Alzheimer’s disease. mented-reality system to aid cognition when combined with Due to the sharp decline in his mental acuity, he is often un- software for scene interpretation, facial recognition, context able to remember the names of friends and relatives; he also awareness, and voice synthesis. When Ron looks at a person frequently forgets to do simple daily tasks. He faces an uncer- for a few seconds, that person’s name is whispered in his ear tain future that’s clouded by a lack of close family nearby and along with additional cues to guide Ron’s greeting and inter- limited financial resources for professional caregivers. Even actions; when he looks at his thirsty houseplant, “water me” is modest improvements in his cognitive ability would greatly whispered; when he looks at his long-suffering dog, “take me improve his quality of life, while also reducing the attention out” is whispered. demanded from caregivers. This would allow him to live inde- In this example, low-latency, high-bandwidth wireless access pendently in dignity and comfort for many more years, before to cloudlet resources is an essential ingredient for the “magic he has to move to a nursing home. glasses” to be able to execute computer vision algorithms for scene analysis and facial recognition at real-time speeds. This is Fortunately, a new experimental technology might provide only one of many new applications that we can imagine. to static client and server hardware.1 ware’s capabilities. In the lab and with it doesn’t take a giant leap of faith to At any given cost and level of technol- ample computing resources, the state recognize their future potential. The ogy, considerations such as weight, size, of the art for applications such as face real challenge lies in sustaining their battery life, ergonomics, and heat dis- recognition, speech recognition, and state-of-the-art performance and qual- sipation exact a severe penalty in com- language translation is near-human ity in the wild—under highly variable putational resources such as processor in performance and quality. As Fig- conditions on lightweight, energy- speed, memory size, and disk capacity. ure 1a shows, for example, researchers efficient, resource-impoverished mobile From the user’s viewpoint, a mobile de- achieved Spanish-English translation hardware. vice can never be too small or light or comparable to human quality in 2006 have too long a battery life. Although on a 100-node computing engine by us- The Limits of Cloud Computing mobile hardware continues to evolve ing large online corpora and a context- An obvious solution to mobile devices’ and improve, it will always be resource- based machine translation algorithm.2 resource poverty is to leverage cloud poor relative to static hardware—sim- For the IBM BLEU metric used in the computing. A mobile device could ex- ply put, for the hardware that people figure, scores above 0.7 enter the bilin- ecute a resource-intensive application carry or wear for extended periods of gual human translator range and those on a distant high-performance com- time, improving size, weight, and bat- above 0.8 approach the experienced pute server or compute cluster and sup- tery life are higher priorities than en- professional human translator range. port thin-client user interactions with hancing compute power. This isn’t Face recognition using computer vision the application over the Internet. Un- just a temporary limitation of current is another area in which rapid progress fortunately, long WAN latencies are a technology but is intrinsic to mobility. has occurred over the past decade. Fig- fundamental obstacle. Computation on mobile devices will ure 1b, adapted from Andy Adler and thus always involve a compromise. Michael Schucker’s 2007 comparison Why Latency Hurts Resource poverty is a major ob- of human and automatic face recogni- WAN delays in the critical path of stacle for many applications with the tion performance,3 shows that comput- user interaction can hurt usability by potential to seamlessly augment hu- ers and humans are comparable in this degrading the crispness of system re- man cognition because such applica- task today. Although several technical sponse. Even trivial user–application tions typically require processing and improvements for practical deployment interactions incur delays in cloud com- energy that far outstrips mobile hard- are still needed in such applications, puting. Humans are acutely sensitive to OCTOBER–DECEMBER 2009 PERVASIVE computing 3 VIRTUAL MACHINES 0.85 0.8 Year Computer Computer Indeterminate Worse/ Human scoring range worse better (%) Better than than 0.7 human (%) human (%) 1999 87.5 4.2 8.3 21.0 0.6 2001 87.5 8.3 4.2 10.5 0.7447 0.5 0.7289 Bleu scores 2003 45.8 16.7 37.5 2.75 0.5551 0.5610 0.4 0.5137 2005 37.5 33.3 29.2 1.13 0.3859 0.3 Google Google Systran SDL Google CBMT 2006 29.2 37.5 33.3 0.78 Chinese Arabic Spanish Spanish Spanish Spanish (‘06 NIST) (‘05 NIST) ’08 top lang (a) Based on same Spanish test set (b) Figure 1. Near-human quality of cognitive augmentation applications today. Machines are much more capable of matching humans in (a) language translation2 and (b) facial recognition3 than in the past. 100 Min Mean Max Lower bound 90 Thin Thick 80 Berkeley–Canberra 174.0 174.7 176.0 79.9 70 Berkeley–New York 85.0 85.0 85.0 27.4 60 Berkeley–Trondheim 197.0 197.0 197.0 55.6 50 CDF Pittsburgh–Ottawa 44.0 44.1 62.0 4.3 40 Pittsburgh–Hong Kong 217.0 223.1 393.0 85.9 30 Thin 100ms Pittsburgh–Dublin 115.0 115.7 116.0 42.0 Thin 66ms 20 Pittsburgh–Seattle 83.0 83.9 84.0 22.9 Thin 33ms 10 Thick 0 0 10 20 30 40 50 60 70 80 90 (a) Smoothness (frames per second) (b) Network latency hurts interactive performance even with good bandwidth: (a) a highly interactive visualization application’s measured output frame rate under two different configurations: “Thick” (a local machine with hardware graphics acceleration) and “Thin” (a remote compute server on a 100 Mb/s network, with round trip times ranging from 33ms to 100ms) (b) measured Internet2 round trip times between representative sites confirm that the 33-100ms range lies well within the range of observed latencies in the real world.