Method for operating a network and a corresponding network
A method for operating a network is provided, where an occupation level of at least one switch queue of at least one network switch is estimated. Data regarding an association between at least one path delay and a corresponding switch queue occupation level of the at least one switch queue is provided or collected. The data is fed to a machine learning model associated with the at least one switch queue. The machine learning model is trained on the basis of the at least one path delay or the data to predict a switch queue occupation level of the at least one switch queue. Information resulting from the trained machine learning model or the predicted switch queue occupation level is used for making a real-time traffic steering decision for load balancing between network paths.
1 . A method for operating a network, wherein an occupation level of at least one switch queue of at least one network switch is estimated, comprising the following steps:
providing or collecting data regarding an association between at least one path delay and a corresponding switch queue occupation level of the at least one switch queue, wherein the data includes path delays and/or switch queue occupations or switch queue occupation levels that are quantized by reducing a precision according to a statistical distribution;
feeding the data to a machine learning model associated with the at least one switch queue;
training the machine learning model on a basis of the at least one path delay or the data to predict a switch queue occupation level of the at least one switch queue; and
using information resulting from the trained machine learning model or using the predicted switch queue occupation level for making a real-time traffic steering decision for load balancing between network paths, wherein the trained machine learning model is disposed at an assigned candidate executor host in the network.
2 . The method according to claim 1 , wherein the collected data results from historical measurements from the network and/or simulated data.
3 . The method according to claim 1 , wherein the at least one path delay is a one-way path delay regarding incoming traffic to a respective network switch.
4 . The method according to claim 1 , wherein one or more subsets of network queues and/or corresponding paths are assigned to the candidate executor host or to a set of candidate executor hosts.
5 . The method according to claim 1 , wherein a size of one or more machine learning models is reduced or pruned.
6 . The method according to claim 1 , wherein the assigned candidate executor host is disposed at an end-point or edge of the network.
7 . The method according to claim 1 , wherein different trained machine learning models are distributed to different assigned candidate executor hosts, and wherein each of the different assigned candidate executor hosts monitors a different definable partition of the network.
8 . The method according to claim 7 , wherein partitioning and/or distributing of the different assigned candidate executor hosts is selected to minimize a computation time and/or maximize estimation accuracy.
9 . The method according to claim 1 , wherein at least one explicit congestion notification is sent in a network data plane.
10 . The method according to claim 1 , wherein the data and the machine learning model are split according to network topology and/or to topological information about candidate executor hosts and/or an input data collection process.
11 . The method according to claim 1 , wherein a simulated network for performing packet-level simulation is created.
12 . The method according to claim 1 , wherein the prediction of the switch queue occupation level is used by a control plane or by a data plane.
13 . The method according to claim 4 , wherein a network topology and/or a maximum number of candidate executor hosts is given or provided.
14 . The method according to claim 1 , wherein the path delays of the network are quantized according to the statistical distribution, and wherein the switch queue occupations of the network are quantized according to a number of occupation levels.
15 . The method according to claim 5 , wherein the size of one or more machine learning models is reduced or pruned by reducing a number of path delays.
16 . The method according to claim 6 , wherein one or more devices are executor hosts, and wherein the one or more devices generate at least one explicit congestion notification.
17 . The method according to claim 9 , wherein the at least one explicit congestion notification is sent in the network data plane to enable in-band reconfiguration of packet forwarding operations.
18 . A method for operating a network, wherein an occupation level of at least one switch queue of at least one network switch is estimated, comprising the following steps:
providing or collecting data regarding an association between at least one path delay and a corresponding switch queue occupation level of the at least one switch queue;
feeding the data to a one or more machine learning models associated with the at least one switch queue;
training the one or more machine learning models on a basis of the at least one path delay or the data to predict a switch queue occupation level of the at least one switch queue; and
using information resulting from the one or more machine learning models or using the predicted switch queue occupation level for making a real-time traffic steering decision for load balancing between network paths, wherein the one or more machine learning models are disposed at an assigned candidate executor host in the network, wherein the one or more machine learning models are binarized or quantized to transform weights and activations of the one or more machine learning models from real numbers to numbers in a {0, 1} set.
19 . A system comprising one or more hardware processors which, alone or in combination, are configured to provide for execution of a method comprising:
providing or collecting data regarding an association between at least one path delay and a corresponding switch queue occupation level of the at least one switch queue, wherein the data includes path delays and/or switch queue occupations or switch queue occupation levels that are quantized by reducing a precision according to a statistical distribution;
feeding the data to a machine learning model associated with the at least one switch queue;
training the machine learning model on a basis of the at least one path delay or the data to predict a switch queue occupation level of the at least one switch queue; and
using information resulting from the trained machine learning model or using the predicted switch queue occupation level for making a real-time traffic steering decision for load balancing between network paths, wherein the trained machine learning model is disposed at an assigned candidate executor host in the network.