Abstract:

Control of switching fabrics is a complex issue. Depending on the chosen algorithm, a switching fabric exhibits different combinatorial properties that impact its ability to provide services with a given Quality of Service (QoS). The complexity of the algorithm varies depending on the type of switching fabric and the desired properties (e.g., non-blocking in the narrow or wide sense, repackable or rearrangeable fabrics, or blocking fabrics). Algorithms also differ in how they affect the state of the switching fabric (and already established connections) when handling a new request.This article attempts to leverage artificial intelligence to control a switching fabric. Standard machine learning mechanisms were utilized for this purpose. The data used to train the models was derived from simulations of switching fabric operation controlled by known path selection algorithms. While for non-blocking fabrics, the AI model demonstrated connection losses (as expected, since these fabrics are designed to be lossless for specific path selection algorithms), for blocking fabrics, the AI model significantly outperformed traditional approaches in accepting new requests.