Method and system for server assignment using predicted network metrics
A method for assigning a server to provide a resource to a client in a distributed network includes receiving a request for the resource from the client. A network metric is measured at different points in the network. The network metric measurements are input to a deep learning model. Using the model, the network metric is predicted between the client and each of a plurality candidate servers which have the resource and have not had a prior connection with the client. One of the candidate servers is assigned to provide the resource to the client based on the predictions of the network metric.
1. A method for assigning a server to provide a resource to a client in a distributed network, the method comprising:
receiving a request for the resource from the client;
measuring a network metric at different points in the network;
inputting the network metric measurements to a deep learning model;
predicting, using the model, the network metric between the client and each of a plurality candidate servers which have the resource and have not had a prior connection with the client; and
assigning one of the candidate servers to provide the resource to the client based on the predictions of the network metric.
2. The method according to claim 1 , further comprising storing the measurements of the network metric in a measurement database, updating the measurement database as conditions in the network evolve and using the updated measurement database to train the model.
3. The method according to claim 2 , wherein the deep learning model includes latent representations of the client and the servers and a regression model having embeddings and weights which are trained using the updated measurement database according to stochastic gradient descent and backpropagation methods.
4. The method according to claim 3 , further comprising using the latent representations to produce a visualization of the network with respect to the network metric.
5. The method according to claim 1 , wherein the resource is video and the network metric is jitter.
6. The method according to claim 1 , wherein the resource is an audio conference and the network metric is delay.
7. The method according to claim 1 , further comprising providing the resource to the client from the assigned server.
8. The method according to claim 1 , wherein a type of the network metric which is measured and used for the prediction is based on a type of the resource which is requested.
9. A system for assigning a server to provide a resource to a client in a distributed network, the system comprising:
a client-neural network interface connected to the client and operable to receive a request for the resource from the client;
a measurement database containing measurements of a network metric; and
a neural network component operable to predict, using a deep learning model in which the measurements of the network metric are input, the network metric between the client and each of a plurality of candidate servers which have the resource and have not had a prior connection with the client,
wherein the client-neural network interface is operable to assign one of the candidate servers to provide the resource to the client based on the predictions of the network metric.
10. The system according to claim 9 , wherein the client-neural network interface is operable to update the measurement database as conditions in the network evolve.
11. The system according to claim 10 , wherein the deep learning model includes latent representations of the client and the servers and a regression model having embeddings and weights, the neural network component being operable to train the embeddings and weights using the updated measurement database according to stochastic gradient descent and backpropagation methods.
12. The system according to claim 9 , wherein the client-neural network interface is operable to determine different types of the network metric to be used based on a type of the resource requested by the client.
13. The system according to claim 12 , wherein the client-neural network interface is operable to determine that the network metric is jitter based on the resource being video.
14. The system according to claim 12 , wherein the client-neural network interface is operable to determine that the network metric is delay based on the resource being an audio conference.
15. The system according to claim 9 , wherein the measurement database includes measurements for a plurality of network metrics including jitter and delay, and wherein the client-neural network interface is operable to select one of the network metrics to be used by the neural network component for the predictions.
16. The method according to claim 1 , wherein the distributed network is a content delivery network (CDN), cloud computing network, or a peer to peer (P2P) network in which the candidate servers which host the resource are only a part of a total number of servers in the distributed network.
17. The method according to claim 1 , wherein the measurements of the network metric at the different points of the network which are used as the input to the deep learning model are taken continuously or regularly.
18. The method according to claim 1 , wherein client-server pairs in the distributed network are represented by vectors and fed into a regression model whose parameters are learned by the deep learning model to predict the network metric for each of the client-server pairs.