Abstract: A method and system are described for automatically determining network utilization bounds in order to meet per-link packet loss rate targets through the network. A loss-verses-utilization model is continuously re-generated as part of a feedback loop to update link utilization targets. A target loss rate can be used to generate a link utilization target value using the model. The link utilization target value can be used in a variety of network update models, such as traffic engineering, capacity planning and risk management. Updates can be applied to individual network links, network devices as a whole, or groups of devices (based on type or position in the network).

Abstract: An optimization agent optimizes configuration parameters of a shared memory buffer for ports of a network device (e.g., router). The optimization agent located on the network device (or remotely) takes an input state from the network device's switching logic, such as the packets transmitted or dropped per port, and determines port buffer settings to optimize the settings to achieve a goal. The buffer settings can include minimum and maximum buffer limits per port and the goal can be to maximize port utilization or minimize port loss. The optimization agent can then take an action by applying the buffer configurations to the switching logic and repeat the process iteratively until the goal is reached.

Abstract: Anomaly detection for one or more streams of time-series data can use an encoder/decoder pair, such as in a variational autoencoder (VAE) in combination with an aggregator or classifier, such as a random isolation forest (RIF). A particular application relates to detecting anomalies in network updates in a large number of network devices that can transmit the updates to a collector for analysis. The encoder/decoder pair can include a neural network with long short-term memory cells or similar type cells. Using the combination, a single anomaly score can be produced from multiple streams of the time-series data.