Edge Computing has multiple facets, there is Edge Computing, Fog Computing, MEC (Multi-Access Edge Computing); but the general concept of “Edge” is about placing the workload closer to the data source / service consumer.
Traditionally, users would request services where the application would fetch data -like a video- in a specific server that could be hundreds of kilometres away, maybe on another continent, or ask a cloud provider for a service like Siri requiring quick and heavy computation; as traffic increased this became unfeasible as it would create network congestion and result in poor latency: if many users in America requested data in European servers and vice-versa, the cables between the two would have to deal with too much traffic.
Looking at the type of data that puts the most pressure on the networking infrastructure, we can suppose -rightly so- that video takes the biggest global traffic share; in the same way that photos and videos take most of the space on our phones. Sending all those pixels requires a lot of bandwidth: The latest report on Internet traffic by Sandvine shows video streaming in the first position with 60% of the global internet traffic in download and 22% in upload!
As video streaming became more prominent, and companies such as Netflix wanted to give a better user experience with less latency (Netflix alone represents 13% of the global internet traffic), they turned to CDNs (Content Delivery Networks) where Netflix would put their shows on many servers, throughout a country in order to be closer to the consumer and reduce the load on the overall infrastructure. Additionally, to avoid copying their whole library of shows, they used popularity models based on usage to put the correct shows in the correct servers close to the users watching them. All this effort of building their own CDN enabled Netflix to offer unrivalled user experience with very low latency and saved money for the ISPs (Internet Service Providers) who could connect Netflix’s CDN directly into their network instead of having to scale their whole internet infrastructure to deal with an increased streaming demand.
Now as we turn towards 5G and ARM-based processors, we can see two distinct trends:
While CDNs don’t compute -they just store content- we have another use case for bringing computation closer to the edge, essentially creating smaller, distributed datacentres / servers that would help offload the centralized servers and reduce latency with end users. For example, Siri wouldn’t be possible without smaller datacentres on the edge receiving the audio, translating the speech to text, running natural language processing to understand the question and seek its answer using an API to connect to the weather channel before returning the temperature in a few milliseconds… It seems almost magical to get an answer so quickly because if the request had to travel back to a centralized server, it would take several seconds before getting a result. In so doing, smaller data centres on the edge not only help offload the network and the centralized servers, but also enable new solutions previously unfeasible.
Similarly, one could argue the new 5G architecture based on Open RAN is itself a form of Edge Computing, given that it’s disaggregating computation from a BBU (Baseband Unit) into a CU and DU (Centralised Unit and Distributed Unit); essentially pushing the computation towards the edge, closer to the antenna (Radio Unit). Although different from the type of computation talked about previously, it also reduces the cost for telecom operators and optimises their network topology by offloading the traffic from some of their centralized units.
Although video streaming is different from live video feeds from surveillance cameras, it helps us understand the throughput pressure it can have on the infrastructure. There are more than 1 billion surveillance cameras in the world right now and it is estimated to double by 2024. With the new generation of cameras that are rolling out: 1 billion cameras filming in 4K @ 25fps using the high performance HEVC (H.265 codec) would have a constant stream of 2.5 PB/s (that’s 2.5 Petabytes per second). The current global internet capacity is 600 Tbps (75 TB/s); while this doesn’t take into account private/local networks, it shows we need to take drastic measures to reduce this future load on our network infrastructure.
Specifically, with 500+ million cameras, China had a problem with network congestion. Not only did they struggle keeping up with data storage and computation when building servers; but the biggest bottleneck was the network itself: the throughput was growing too rapidly compared to what the cables and routers could handle. Rather than sending the whole video feed from the camera to the datacentres to be processed; we could use on-device processing to bring the computation closer to the Edge and significantly reduce the data being transferred.
With the renewed popularity of ARM as the processors became powerful enough to rival Intel and x86 -while keeping breathtaking energy efficiency- new possibilities arose: enabling on-device processing for computation-heavy workloads like Computer Vision. This means we can add a powerful chip without significantly changing a smart camera’s design and power consumption.
For example: highway traffic cameras could run Computer Vision algorithms that would detect cars or highway activity, measuring traffic flow and congestion. Instead of sending all the pixels of each frame, the camera could send a very lightweight string of data describing the relevant information in the frame: “a red Nissan with plate number ‘GB21 XHB’ is going 60 mph” or “the congestion index is 0.83”. Furthermore, if the camera detects something abnormal, like someone walking on the road (or what the authorities are specifically looking for) it can keep sending the original stream of video back to the data centre for further analysis and storage.
Some existing solutions are sending the metadata from the on-device analysis in sync with the video feed, for the servers to have an easier task and help them “look for” specific moments. In the future, once Computer Vision algorithms reach a high-enough success rate, (or do simple enough tasks) we can delete the video feed altogether and save all the throughput.
In this article we talked about different ways Edge Computing can help reduce network congestion by using CDNs (Content Delivery Networks), distributed datacentres and On-device processing.
If you wish to know more about Edge Computing and how Net Reply can help you with Edge solutions, contact Paul Beglin.