Towards In-Network Industrial Feedback Control

Abstract

Controlling physical machinery and processes is at the core of production automation. However, challenged by inflexibility, automation and control is evaluating to outsource this control to resourceful cloud environments. While this enables to derive better control through a plethora of measurements, it challenges the control quality through delay introduced through networks. In this paper, we show how to unify control and communication by offloading delay sensitive control tasks from the cloud to local network elements-a previously unexplored area for in-network processing-enabling both, ultra high quality-of-control while enabling scalable orchestration through cloud environments. Our implementation demonstrates how we combine state of the art control with communication. We achieve this by expressing the control and the datapath in P4 which we synthesize to BPF programs that we execute in XDP environments on Netronome SmartNICs. Further, we highlight the demands of control towards communication to build more involved and complex in-network controllers.

Control-over-the-cloud: A performance study for cloud-native, critical control systems

“…Research on deploying traditional feedback control loops to a cloud in general, is often focused on making cloud computing platforms and intermediate networks behave as real-time systems. Targeted efforts have, for example, been made in the areas of deterministic dynamic networks [5], resource allocation in data centers [18], and feedback control implemented in the network [12]. Nevertheless, successful attempts at deploying feedback control loops that span a cloud Virtual Machine (VM) and back, were first made relatively recently [4].…”

A Case for Integrated Data Processing in Large-Scale Cyber-Physical Systems

“…Following a different approach, the expressiveness of programs that can be executed on network devices can be further increased: By offloading so-called Extended Berkeley Packet Filter (eBPF) programs, originally developed for fast handling of network packets in the kernel [36], to network hardware, control algorithms can be run within the network [37,38]. This leads to dramatic increases in the control quality under deteriorating latency and jitter conditions.…”

Towards Executing Computer Vision Functionality on Programmable Network Devices

“…It, e.g., neither defines floatingpoint operations, nor the amount of individual operations an action must at least contain and not even the number of available MAU stages. More complex processing may hence need to be reformulated according to the available functionality [23]. Some targets may offer the possibility to re-inject packets at the beginning of the pipeline (recirculation), allowing it to pass through the same MAUs multiple times, which enables the implementation of longer processing procedures.…”

“…By partially offloading processing steps to network elements and thus avoiding overheads introduced by multi-hop paths and network stacks of operating systems, the applicability of In-Network Processing (INP) has already been shown to be beneficial for network management-related applications such as load balancing [20] or the detection of heavy hitters [29] and of DDoS attacks [16]. The possibility to reduce the overall traffic in map-reduce and machine learning applications via in-network data aggregation has been studied in [24,25], as has the reduction of latency in industrial feedback control by offloading simple control algorithms to emulated programmable switches [23]. Further work concerns e.g., consensus protocols [18] and key-value caching [11].…”