Forecasting SDN fabric saturation and machine learning-based flow admission control

A tangible, non-transitory, computer-readable medium storing program instructions that cause a device of a software defined wide area network (SD-WAN) to execute a process comprising:

• predicting, by the device, characteristics of a new traffic flow to be admitted to the SD-WAN, based on a set of initial packets of the new traffic flow;
• predicting, by the device, an impact of admitting the new traffic flow to the SD-WAN, based in part on extrinsic data or exogenous information regarding the SD-WAN;
• admitting, by the device, the new traffic flow to the SD-WAN, based on the impact predicted by the device;
• and
• using, by the device, reinforcement learning to adjust one or more call admission control (CAC) parameters of the SD-WAN, based on captured telemetry data regarding the new traffic flow admitted to the SD-WAN,
• wherein the extrinsic data or exogenous information regarding the SD-WAN comprises at least one of:
• 1) external table lookup information indicative of routing table information or one or more tunnel destinations in the SD-WAN, 2) information regarding a layout of virtual machines (VMs) or containers in which an application is executed, an application load, or transport ports associated with the application, or 3) data indicative of application types associated with existing traffic in the SD-WAN.

An apparatus, comprising:

• one or more network interfaces;
• a processor that is coupled to the one or more network interfaces;
• and
• a memory configured to store a process that is executable by the processor, the process when executed configured to:
• predict characteristics of a new traffic flow to be admitted to a software defined wide area network (SD-WAN), based on a set of initial packets of the new traffic flow;
• predict an impact of admitting the new traffic flow to the SD-WAN, based in part on extrinsic data or exogenous information regarding the SD-WAN;
• admit the new traffic flow to the SD-WAN, based on the impact predicted by the apparatus;
• and
• use reinforcement learning to adjust one or more call admission control (CAC) parameters of the SD-WAN, based on captured telemetry data regarding the new traffic flow admitted to the SD-WAN,
• wherein the extrinsic data or exogenous information regarding the SD-WAN comprises at least one of:
• 1) external table lookup information indicative of routing table information or one or more tunnel destinations in the SD-WAN, 2) information regarding a layout of virtual machines (VMs) or containers in which an application is executed, an application load, or transport ports associated with the application, or 3) data indicative of application types associated with existing traffic in the SD-WAN.