Investigating the Applicability of In-Network Computing to Industrial Scenarios
Ike Kunze, René Glebke, Jan Scheiper, Matthias Bodenbenner, Robert Schmitt, Klaus Wehrle
Abstract
Moving computation functionality into the network and onto networking devices has recently shown great promise for performance improvements, e.g., by reducing path lengths and processing latencies, or by increasing bandwidth efficiency. Yet, In-Network Computing (INC) is challenged by the limitations of networking hardware which is generally not designed for complex calculations. Thus, while possible, developing INC approaches is a constant struggle between desired and available functionality. In this paper, we demonstrate the applicability of INC to industrial contexts by offloading a seemingly simple task from an industrial assembly scenario - coordinate transformations - to programmable switches and SmartNICs. We find that even such a task puts heavy demands on the devices, but that the right dose of approximation and diligent problem reformulation can enable the necessary operations on all platforms, thus setting the stage for improving latencies by significantly reducing communication paths. Yet, our results further indicate that these gains in latency come in hand with a trade-off regarding the achievable accuracy.