Method and system for training a model for resource deployment in a distributed compute network
There is provided a computer-implemented method of reinforcement learning for training a machine learning model configured to provision and/or deprovision virtual application instances of an application in a distributed compute network having one or more compute nodes, the method being performed by at least one hardware processor and the method comprising: a) defining an action space for a training environment representative of the distributed compute network, the training environment having a plurality of states and, for each state, one or more available actions in the action space, each action being operable to transition the distributed compute network in the training environment from a current state to a new state; b) selecting one or more available actions in the action space to define a new state; c) determining a reward for the one or more selected actions based upon a predicted change in load characteristics for the distributed compute network in the training environment in moving from the current state to the new state, the predicted change in load characteristics being determined at least in part based on training data relating to the distributed compute network, wherein the training data comprising simulated load data relating to one or more parameters of one or more compute nodes; and d) training the machine learning model using the predicted change in load characteristics and the reward values to generate an updated trained machine learning model. nodes for provisioning of virtual application instances of the application.
1 . A computer-implemented method of provisioning resources in a distributed compute network comprising one or more routing nodes and one or more compute nodes configured to host one or more virtual application instances of an application thereon, the method being performed by at least one hardware processor and comprising:
(a) receiving, by a system manager, resource load data relating to the computational loads resulting from one or more virtual application instances running on one or more compute nodes corresponding to a current state of the distributed compute network;
(b) determining, by the system manager, whether the values of the resource load data corresponding to the current state of the distributed compute network fall outside one or more predetermined target ranges for the resource load data and, if so determined, performing the step of:
(c) dynamically performing, utilizing a trained machine learning model, provisioning of virtual application instances and/or deprovisioning of virtual application instances on one or more compute nodes to generate a new state of the distributed computing network,
wherein the trained machine learning model is trained by a reinforcement learning training method comprising
(d) defining an action space for a training environment representative of the distributed compute network, the training environment having a plurality of states and, for each state, one or more available actions in the action space, each action being operable to transition the distributed compute network in the training environment from a current state to a new state,
(e) selecting one or more available actions in the action space to define a new state,
(f) determining a reward for the one or more selected actions based upon a predicted change in load characteristics for the distributed compute network in the training environment in moving from the current state to the new state, the predicted change in load characteristics being determined at least in part based on training data relating to the distributed compute network, wherein the training data comprising simulated load data relating to one or more parameters of one or more compute nodes, and
(g) training the machine learning model using the predicted change in load characteristics and the reward values to generate an updated trained machine learning model.
2 . The computer-implemented method according to claim 1 , wherein the training method further comprises:
(h) generating training data comprising simulated load data relating to one or more parameters of one or more compute nodes.
3 . The computer-implemented method according to claim 2 , wherein step (h) comprises generating training data by generating a time-series sequence of simulated values representative of one or more parameters of one or more compute nodes.
4 . The computer-implemented method according to claim 3 , wherein the time-series sequence of simulated values have pre-defined ranges.
5 . The computer-implemented method according to claim 4 , wherein the time-series sequence of simulated values for the one or more parameters of one or more compute nodes is generated using one or more mathematical functions.
6 . The computer-implemented method according to claim 5 , wherein the one or more mathematical functions are selected from the group of: a constant function; a linear function; a quadratic function; and a polynomial function.
7 . The computer-implemented method according to claim 1 , wherein the one or more parameters are selected from one or more of: volume of requests per unit time; process count; memory footprint; number of virtualization containers; disk I/O; network and port usage; and percentage of CPU utilization for one or more servers of one or more compute nodes.
8 . The computer-implemented method according to claim 1 , wherein step (f) comprises utilizing a load predictor model trained on the training data to predict the change in load distribution on the distributed compute network resulting from one or more actions in step (e).
9 . The computer-implemented method according to claim 8 , wherein, prior to step (e), the method comprises:
(i) training the load predictor model on the training data.
10 . The computer-implemented method according to claim 8 , wherein the load predictor model comprises a machine learning model.
11 . The computer-implemented method according to claim 1 , wherein the available actions in step (e) comprise one or more of: provision one or more virtual application instances of an application on one or more compute nodes; deprovision one or more virtual application instances of an application on one or more compute nodes; and take no action.
12 . The computer-implemented method according to claim 11 , wherein step (f) is performed after the actions have been implemented in step (e) to define the new state.
13 . The computer-implemented method according to claim 1 , wherein, subsequent to step (f), the method further comprises:
(j) determining whether a reward target has been met and, if so, progressing to step (g).
14 . The computer-implemented method according to claim 13 , wherein, if the reward target is not met at step j, the method further comprises:
(k) iteratively repeating step (e) and step (f) until the reward target is met.