Which problems with cloud computing make edge computing useful?

Question

I recently read an article in The Register, Don't let cloud slurp all your data. Chew it on the edge, says HPE:

The basic pitch is that HPE's gear can do compute on your shop floor without taking up large chunks of your floorspace, meaning you don't need to splash out on collecting and moving data back and forth, or spend megabucks on cloud services.

However, edge computing (which essentially just seems like a buzzword for locally processing data!) seems to have a few problems with it to me. Wikipedia quote a source which says, "Cloud computing is cheaper because of economics of scale", so surely edge computing misses out on the benefits you get from massive-scale computing in data centres?

Why would edge computing be useful in some cases? Is it only really useful in cases where huge amounts of data need to be sent, where it would be impractical to send it over the Internet?

Answer 1

Remember, you are always processing data at the edge, even if it's not obvious. The choice to sample data at a particular frequency, whether that is 1Hz or 100Khz (particularly with analogue data), is a form of edge processing. Very few scenarios will transmit data at the maximum clock cycle of the processor.

Some scenarios where explicit edge processing is useful

• Bandwidth constraints. The often-quoted example is in the oil and gas industry where a lot of telemetry is gathered, and the bandwidth from the remote location is limited.
• Low latency requirement. Especially in control systems where something needs to happen in response to data. When sub-second responses mean the difference between normal operation and failure, a second or two round-trip to the cloud is unacceptable.
• Extremely low value of data. If nothing is likely to be gained from recording data, even in the future with sophisticated analytics, you might as well summarise and drop it at the edge. For example, mains power in the EU is expressed as 'nominally 230 V ±10% at 50Hz'. Given that you can handle that range, there is no point in transmitting the precise voltage every 100ms. If is better to transmit 'from this time to this time, the voltage was nominal'.
• Ordering of events is important. Devices have a narrower scope and can perform some derivations easier than the cloud. One of these is where the derivation requires that the data is in order. For example, it is easy on a device when measuring, say voltage, when it enters the on or off state. Timestamping the voltage measurement and sending it to the cloud makes the derivation harder as data may come out of order, delayed or be completely lost.
• As a failure mode. Connectivity to the cloud, or at least some of the processing, is likely to fail. All systems should be able to perform some edge-processing, even if it is dumbed-down, as a failure mode.
• Non-IP mesh The cloud will only be accessible via a gateway for non-IP devices, such as those that communicate over a low-power network or CAN bus-type wired connection. It may be preferable for devices communicate directly with each other in a 'mesh', in which case IP connectivity isn't needed, never mind cloud connectivity. Mesh is the holy grail of edge computing, but difficult to do.
• Privacy, trust, and regulatory issues Data can be collected that may be useful and valuable for analytics, but getting customer consent to collection and storage of the data may be difficult, and the cost and risk associated with managing access to the data may be too high. Facial recognition, and most image processing, is the obvious example. Another example is that all modern vehicles have GPS and connectivity built-in for emergency call response. Vehicle manufacturers and dealers could derive valuable insights from a stream of GPS data, but location data is personally identifiable in this scenario and shouldn't be collected, transmitted, processed and stored.
• Security We are aware that IoT is currently suffering from security issues. While it is possible to create a secure environment with cloud-connected devices, security is a major issue that needs to be actively worked on. Highly secure gateways, edge-based threat detection and other as-yet unknown defense mechanisms are going to be necessary in most environments.

When considering material about edge processing, remember to consider the source. Existing IT vendors that haven't cracked the cloud market (IBM, HP, Cisco) are worried that IoT, as the next big thing, bypasses them altogether. As a result they will aggressively market edge processing. Indeed, 'fog computing' is a term created by Cisco to have a 'cloud' closer to the ground (or something). Obviously cloud vendors are marketing the reverse, for their own bottom line.

Answer 2

A security camera would be the obvious example. if you want to do motion detection on the image (or face recognition), you have the choice of either streaming the video feed to the cloud, or processing it locally. This sort of application will probably be marginal for upstream bandwidth on a lot of domestic internet connections today - that gives more incentive to perform either processing or at least compression on the edge.

The audio decoding of a smart home hub is another example. Certainly, the wake-word detection makes sense to perform at the hub (if only for privacy). Once the edge node has the processing capability required for voice control, maybe it makes sense to push that processing to the edge. For the service provider, this is a free gain - the node already needs a low-end linux capable platform, so why not make use of this resource rather than providing bandwidth and CPU in the cloud. Significantly, the bandwidth is likely to cost as much as the CPU in this context.

Even if the bandwidth load is reduced, an application may still be able to use edge computing and save cloud resource by making use of compute power which already exists in the edge node. Remember, the edge needs to encrypt any data it upstreams - this may cost as much as processing it locally.

What has changed in the equation is the cost of provisioning an additional number of DMIPS at the edge (either in home, in the cell-site or in the backhaul). This isn't so much a problem with the cloud, but it changes the cost equations.