How We Created Edge Computing Edge Computing Processes Data on Infrastructure That Is Located Close to the Point of Data Creation
Total Page：16
File Type：pdf, Size：1020Kb
reverse engineering How we created edge computing Edge computing processes data on infrastructure that is located close to the point of data creation. Mahadev Satyanarayanan recounts how recognition of the potential limitations of centralized, cloud-based processing led to this new approach to computing. Mahadev Satyanarayanan n 2009, together with Victor Bahl, Ramón University proposed a technique that is now Client Cáceres and Nigel Davies, I published a widely used for many compute-intensive Viceroy Remote Speech Odyssey RPC paper entitled ‘The case for VM-based applications: a mobile device offloads heavy Janus front-end I API server cloudlets in mobile computing’. Today, the computations over a wireless network to Speech work is generally viewed as the founding a server that is much more powerful than warden manifesto of edge computing. At a time the mobile device. The approach was first RPC when there was widespread euphoria demonstrated in the Odyssey system about the limitless possibilities of cloud (Fig. 1) in 1997. An important aspect of Local computing, the paper put forward an this implementation was Odyssey’s ability Janus alternative viewpoint. It argued that the to select the optimal execution mode server extreme consolidation (the concentration (local, remote or hybrid) based on runtime of computing resources into a few large factors such as current network bandwidth. data centres) implicit in cloud computing Odyssey was thus the technical forerunner Fig. 1 | Offloading from a mobile device. The Janus would fundamentally limit its ability to of today’s mobile speech-to-text systems, as speech-recognition application was modified to sustain latency-sensitive and bandwidth- well as modern mechanisms for adaptive operate in one of three modes in Odyssey: local, hungry applications that would emerge offload. Since 2001, this approach has also remote and hybrid. Image reproduced from in the future. It also identified bandwidth been known by the term cyber foraging B. D. Noble et al., Proc. 16th ACM Symp. Operating scalability of cloud-based applications based and has been a key area of mobile Systems Principles 276–287 (Association for on video data from sensors as a key concern. computing research. Computing Machinery, 1997). Viceroy is a system To support these future applications, the The emergence of cloud computing module described by B. D. Noble et al. paper argued for a dispersed infrastructure around 2007 both simplified and of micro-datacentres called cloudlets, complicated offloading. On the one hand, which avoids extreme consolidation while the cloud was the natural answer to where preserving cloud computing attributes such offloaded execution should be performed. ‘data centre in a box’ and the concept of a as multi-tenancy with strong isolation (in On the other hand, the likely distances to tiered architecture with the cloud at tier 1, other words, the ability to concurrently run the cloud, necessary for the consolidation cloudlets at tier 2 and cloudlet-associated code from different parties that do not trust implicit in cloud computing, were mobile devices at tier 3. To capture the ideas each other). With the recent emergence of problematic. End-to-end communication that emerged from the meeting, we wrote edge computing, our proposal has become over a wide area network to a distant cloud the 2009 paper. mainstream. But what led us to this idea? involves many network hops and results As we learned over the next few years, In 1993, I wrote a short thought piece on in high round-trip times. Bottlenecks for just writing a paper is, alas, not sufficient to the topic of mobile computing, a field that network bandwidth are also likely. convince many sceptics. It took numerous was just emerging at the time. This, I believe, By 2008, I was convinced that embracing years of empirical measurements, and is the first paper where the inherent resource cloud computing for offloading was a sterile the implementation of applications that poverty of mobile devices was identified strategy: it would never be able to sustain are critically dependent on low latency or as a key long-term constraint of mobile applications such as augmented reality, bandwidth scalability, to overcome scepticism computing. As I wrote in the article: “Mobile which were starting to emerge. I shared about the need for edge computing. elements are resource-poor relative to static these concerns with researchers in mobile By 2018, the seed that we had planted in elements. Regardless of future technological computing including Bahl of Microsoft 2009 had grown into industry-wide interest advances, a mobile unit’s weight, power, Research, Roy Want who was then at Intel, and activity. While edge computing is still in size and ergonomics will always render Cáceres who was then at AT&T Research its infancy, there is no question that it is it less computationally capable than its and Davies of Lancaster University. They here to stay. ❐ static counterpart. While mobile elements agreed with my concerns, and expressed will undoubtedly improve in absolute interest in exploring this topic more deeply. Mahadev Satyanarayanan ability, they will always be at a relative We met for a day and a half of brainstorming School of Computer Science, Carnegie Mellon disadvantage.” In the 25 years or so since in October 2008, hosted by Bahl at the University, Pittsburgh, PA, USA. that article, the prediction has remained Microsoft Research office in Redmond, e-mail: [email protected] consistently true. Washington. Many of the key themes of To overcome this fundamental limitation, edge computing emerged from this meeting, Published online: 16 January 2019 my colleagues and I at Carnegie Mellon including the concept of a cloudlet as a https://doi.org/10.1038/s41928-018-0194-x 42 NATURE ELECTRONICS | VOL 2 | JANUARY 2019 | 42 | www.nature.com/natureelectronics.