# US 2020/0287976 A1 · Adaptive routing of branch traffic in software-defined wide area network (sdwan) deployments

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2020/0287976 A1
- Document ID: US20200287976A1
- Status: Published application
- Kind: A1
- Application number: 16/294,388
- Filed: 2019-03-06
- Published: 2020-09-10
- Earliest claimed: 2019-03-06
- Assignees: Hewlett Packard Enterprise Development LP
- Applicants: Hewlett Packard Enterprise Development LP
- Inventors: Isaac Theogaraj; Manjunatha GS; Sivappirakasam Muthiah
- Web page: https://patentgrep.com/patent/US20200287976A1

## CPC

- H04L 67/51: Discovery or management thereof, e.g. service location protocol [SLP] or web services
- G06F 9/45558: {Hypervisor-specific management and integration aspects}
- G06N 3/0442: characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
- G06N 3/08: Learning methods
- G06N 3/09: Supervised learning
- G06N 20/00: Machine learning
- H04L 12/66: Arrangements for connecting between networks having differing types of switching systems, e.g. gateways
- H04L 41/16: using machine learning or artificial intelligence
- H04L 43/0864: Round trip delays
- H04L 45/302: Route determination based on requested QoS
- H04L 47/28: in relation to timing considerations
- H04L 67/101: based on network conditions
- G06F 2009/45595: {Network integration; Enabling network access in virtual machine instances}
- G06N 3/044: Recurrent networks, e.g. Hopfield networks
- H04L 45/64: using an overlay routing layer
- H04L: TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION (arrangements common to telegraphic and telephonic communication H04M)
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS

## Abstract

A method including selecting, with a network orchestrator, a first virtual internet gateway (VIG) as a primary VIG for a branch gateway, is provided. The method includes selecting a second VIG as a secondary VIG for the branch gateway. The method includes monitoring roundtrip times for multiple packets between the primary VIG and the branch gateway, and between the secondary VIG and the branch gateway, and determining a first forecast roundtrip time associated with the connection between the primary VIG and branch gateway, as well as determining a second forecast roundtrip time associated with the connection between the secondary VIG and the branch gateway. The method includes selecting a new primary VIG based on the first forecast roundtrip time and the second forecast roundtrip time, and further based on a skew of active branches between the primary VIG and the secondary VIG.

## Claims

### Claim 1

A computer-implemented method, comprising:
receiving, at a network orchestrator, a first set of roundtrip times (RTTs) for multiple diagnostic packets transmitted between a branch gateway (BG) and a first virtual internet gateway (VIG) assigned as primary VIG to the BG;
receiving, at the network orchestrator, a second set of RTTs for multiple diagnostic packets transmitted between the BG and a second VIG assigned as secondary VIG to the BG;
determining, based on the first set of RTTs, a first forecast RTT for a packet transmitted between the BG and the first VIG;
determining, based on the second set of RTTs, a second forecast RTT for a packet transmitted between the BG and the second VIG; and
assigning a new primary VIG to the BG based on a smaller of the first forecast RTT and the second forecast RTT and based on a skew of active branches between the first VIG and the second VIG.

### Claim 2 (depends on 1)

The computer-implemented method of claim 1, wherein determining the first forecast RTT comprises:
inputting the first set of RTTs to a recurrent neural network (RNN);
weighting each of the first set of RTTs using respective input weight factors; and
calculating, by the RNN, the first forecast RTT based in part on the weighted first set of RTTs.

### Claim 3 (depends on 2)

The computer-implemented method of claim 2, wherein calculating the first forecast RTT comprises:
transmitting outputs of a set of input neurons of the RNN to a set of intermediate neurons of the RNN;
transmitting outputs of the set of intermediate neurons to a set of output neurons of the RNN; and
outputting the first forecast RTT from the set of output neurons.

### Claim 4 (depends on 3)

The computer-implemented method of claim 3, wherein the first forecast RTT is fed back into at least one of: an input neuron of the set of input neurons and an intermediate neuron of the set of intermediate neurons.

### Claim 5 (depends on 2)

The computer-implemented method of claim 2, wherein the RNN is trained using a measured RTT from a packet sent between the BG and the first VIG in comparison to the first forecast RTT.

### Claim 6 (depends on 1)

The computer-implemented method of claim 1, wherein determining the second forecast RTT comprises:
inputting the second set of RTTs to the RNN;
weighting each of the second set of RTTs using respective input weight factors; and
calculating, by the RNN, the second forecast RTT based in part on the weighted second set of RTTs.

### Claim 7 (depends on 1)

The computer-implemented method of claim 1, wherein the new primary VIG is selected based on the lesser of the first forecast RTT and the second forecast RTT when the skew of active branches between the first VIG and the second VIG is less than a threshold.

### Claim 8 (depends on 1)

The computer-implemented method of claim 1, further comprising:
receiving, at the network orchestrator, a third set of RTTs for multiple diagnostic packets transmitted between a second BG and the second VIG, which is assigned as primary VIG to the second BG;
receiving, at the network orchestrator, a fourth set of RTTs for multiple diagnostic packets transmitted between the second BG and the first VIG, which is assigned as secondary VIG to the second BG;
determining, based on the third set of RTTs, a third forecast RTT for a packet transmitted between the second BG and the second VIG;
determining, based on the fourth set of RTTs, a fourth forecast RTT for a packet transmitted between the second BG and the first VIG; and
assigning a new primary VIG to the second BG based on a smaller of the third forecast RTT and the fourth forecast RTT and based on a skew of active branches between the first VIG and the second VIG.

### Claim 9 (depends on 1)

The computer-implemented method of claim 1, wherein assigning a new primary VIG to the BG comprises sending new redistribution costs to the first VIG and the second VIG such that a local area network including the first and second VIGs prefers to send traffic passing through the BG through the assigned new primary VIG.

### Claim 10 (depends on 1)

The computer-implemented method of claim 1, further comprising assigning a secondary VIG to the BG, and engaging the secondary VIG when a connectivity status of the primary VIG to the BG is lost.

### Claim 11

A network orchestrator, comprising:
a memory storing instructions; and
one or more processors configured to execute the instructions to cause the system to:
receive a first set of roundtrip times (RTTs) for multiple diagnostic packets transmitted between a branch gateway (BG) and a first virtual internet gateway (VIG) assigned as primary VIG to the BG;
receive a second set of RTTs for multiple diagnostic packets transmitted between the BG and a second VIG assigned as secondary VIG to the BG; input the first set of RTTs into a machine learning algorithm;
receive, from the machine learning algorithm, a first forecast RTT for a packet transmitted between the BG and the first VIG;
input the second set of RTTs into the machine learning algorithm;
receive, from the machine learning algorithm, a second forecast RTT for a packet transmitted between the BG and the second VIG; and
assign a new primary VIG to the BG based on a smaller of the first forecast RTT and the second forecast RTT and based on a skew of active branches between the first VIG and the second VIG.

### Claim 12 (depends on 11)

The system of claim 11, wherein the machine learning algorithm is a recurrent neural network (RNN) and the first and second forecast RTTs are outputs of the RNN when the first set of RTTs and the second set of RTTs are inputs of the RNN, respectively.

### Claim 13 (depends on 11)

The system of claim 11, wherein the machine learning algorithm is trained by comparing the first and second forecast RTTs to first and second measured RTTs, respectively.

### Claim 14 (depends on 11)

The system of claim 11, wherein the one or more processors further cause the system to assign new redistribution costs to the first VIG and the second VIG such that infrastructure devices of a local area network including the first VIG and the second VIG prefers to send traffic passing through the BG through the assigned new primary VIG.

### Claim 15 (depends on 14)

The system of claim 14, wherein the new redistribution costs are open shortest path first (OSPF) redistribution costs and the local area network is an OSPF network.

### Claim 16 (depends on 11)

The system of claim 11, wherein the network orchestrator is executed on a cloud device coupled to the BG and the first and second VIGs via the Internet.

### Claim 17

A computer-implemented method comprising:
receiving, at a network orchestrator, a notification that a branch gateway (BG) has joined a network;
assigning a first virtual internet gateway (VIG) as primary VIG for the BG and a second VIG as secondary VIG to the BG;
receiving, at the network orchestrator, a first set of roundtrip times (RTTs) for multiple diagnostic packets transmitted between the BG and the first VIG;
receiving, at the network orchestrator, a second set of RTTs for multiple diagnostic packets transmitted between the BG and the second VIG;
determining, based on the first set of RTTs, a first forecast RTT for a packet transmitted between the BG and the first VIG;
determining, based on the second set of RTTs, a second forecast RTT for a packet transmitted between the BG and the second VIG; and
assigning the second VIG as a new primary VIG to the BG based on the first forecast RTT being larger than the second forecast RTT.

### Claim 18 (depends on 17)

The computer-implemented method of claim 17, wherein the first set of RTTs is received from the first VIG and the second set of RTTs is received from the second VIG.

### Claim 19 (depends on 17)

The computer-implemented method of claim 17, wherein determining the first forecast RTT comprises inputting the first set of RTTs into a machine learning algorithm and receiving the first forecast RTT as an output of the machine learning algorithm.

### Claim 20 (depends on 19)

The computer-implemented method of claim 19, wherein the machine learning algorithm is a recurrent neural network.
