OS strategies for containerized edge applications

Noah Greene, automation product marketing specialist, Phoenix Contact USA: Debian or Ubuntu operating systems with real-time patches are common for edge control applications. Specifically, the PREEMPT_RT patch for Ubuntu allows higher priority tasks to execute immediately, preempting lower priority kernel tasks.

As well, Yocto Project is an open-source toolset for embedded Linux that allows developers to tailor the OS to the hardware. Parts of a Linux distribution that are not important for an edge controller, like a GUI or Office Suite, can be omitted, while things like custom drivers can be added.

What containerization or virtualization technologies, such as Docker or Kubernetes-based frameworks, are supported for deploying applications at the edge? How does the operating system support standard Docker runtimes, and how is the persistent storage handled to prevent SD card or EMMC wear-out from frequent log writes? 

Noah Greene, automation product marketing specialist, Phoenix Contact USA: Docker, Kubernetes, and Podman are container runtimes that I have seen used in the edge control space. I tend to see Docker and Podman at the hardware level, most likely running on a PC close to the machine, while Kubernetes tends to run in the cloud-based environments.

The typical limit for write cycles is around 10,000 for standard SD cards. If no measures are implemented, this can pass rather quickly. Adjusting some settings in the OS and for the container engine/runtime can slow this. For example, adjusting the threshold for when the Linux kernel swaps RAM to disk down from the default will decrease the amount of time data needs to be written to persistent memory, and/or creating temporary storage systems (TMPFS) for volumes that do not need to retain the data between reboots. All the data is stored in RAM instead of on the SD card, limiting the number of write cycles an SD card would go through.

What are the advantages or disadvantages of an open or closed edge environment?

Noah Greene, automation product marketing specialist, Phoenix Contact USA:

Pros of open environments: increased innovation by using community-based collaboration, interoperability between vendors, elimination of vendor lock-in.

Cons of open environments: making sure data is consistent across vendors, ensuring software packages are free of malicious code.

Pros of closed environments: consistent data between nodes, vendor responsible for security updates/patches, integrating nodes can be much smoother.

Cons of closed environments: limited interoperability with third parties, licensing fees can add up over time, legacy systems can lose support.

Why is it important to know which industrial communication protocols an edge controller natively supports and whether additional protocols be added through software or middleware?

Noah Greene, automation product marketing specialist, Phoenix Contact USA: If we think about industrial communication protocols as being spoken/written languages, it is much easier to communicate with someone that speaks your native language. If a system uses Profinet, for example, and the edge controller natively supports Profinet, then connecting it in will be much easier. The data is going to be formatted and transmitted in way that all nodes can understand. On the flip side, if an edge controller natively supports Profinet while the system uses EtherNet/IP, then there will most likely be a need for a protocol converter or custom software to translate the data from one protocol to another. This takes time and experience to get right.

How do compute resources, such as CPU architecture, cores, RAM or storage, affect the ability to run analytics, vision or AI workloads locally?

Noah Greene, automation product marketing specialist, Phoenix Contact USA: Management and allocation of system resources apply to more than just analytics, vision and AI. But in this context, vision and AI consume a lot of resources. While the vision system is doing its thing, analytic tasks that might run concurrently will most likely be slower since it must wait for resources to be freed. Large AI models need more RAM. Edge compute tasks will also need RAM. If the AI workload uses most of the RAM available, then edge compute tasks will have to wait until more RAM is freed up.

Of course, adding additional CPU cores, RAM and storage comes with an increased cost. Manual allocation of resources to tasks and kernel preempting, especially on systems with limited resources, can help keep resources available to the processes when they execute.

Tell us about one of your company’s state-of-the-art product, if any, that involves edge computing.

Noah Greene, automation product marketing specialist, Phoenix Contact USA: The PLCnext Control portfolio is designed with edge control in mind. The hardware PLCs, like the AXC F 2152, combined with what Phoenix Contact calls the Edge Gateway App, make collecting data from the machine using normal PLC I/O and IEC 61131 programming and exchanging this data with cloud services incredibly easy. Even the virtual PLCs can be spun up in an OCI-container, connected to a Profinet fieldbus to gather that data, which can then be shared with cloud services, on-premise databases or whatever is applicable to a given application.