Login| Sign Up| Help| Contact|

Patent Searching and Data


Title:
METRIC BASED CLUSTER SCALING
Document Type and Number:
WIPO Patent Application WO/2019/175641
Kind Code:
A1
Abstract:
Here we have a group of virtual machine nodes in a cluster. There is one virtual machine node called Node Manager which can be one of the virtual machine nodes in the cluster or outside of it. Node Manager periodically aggregates the information for each virtual machine node including CPU load and idle time of the virtual machine node, Memory usage of the virtual machine node, number of connections handled by the virtual machine node, etc. Now based on these parameters Node Manager makes decisions on whether to scale up or down the resources for a particular virtual machine or to add or remove one or more virtual machine nodes in the cluster with the help of empirical data of previously run heavy load tests on the same platform or with the help of machine learning and the dataset of heavy load tests that were run to train it.

Inventors:
SHARMA PRATIK (IN)
Application Number:
PCT/IB2018/051773
Publication Date:
September 19, 2019
Filing Date:
March 16, 2018
Export Citation:
Click for automatic bibliography generation   Help
Assignee:
SHARMA PRATIK (IN)
International Classes:
G06F21/53; G06Q10/00; H04L29/06
Foreign References:
US20150295808A12015-10-15
Download PDF:
Claims:
Claims

Following is the claim for this invention: -

1. In this invention we have a group of virtual machine nodes in a cluster.

There is one virtual machine node called Node Manager which can be one of the virtual machine nodes in the cluster or outside of it. Node Manager periodically aggregates the information for each virtual machine node including Central Processing Unit (CPU) load and idle time of the virtual machine node, Memory usage of the virtual machine node, number of connections handled by the virtual machine node, etc. Now based on these parameters Node Manager makes decisions on whether to scale up or down the resources for a particular virtual machine or to add or remove one or more virtual machine nodes in the cluster. To make sure Node Manager takes appropriate decisions we run heavy load tests on the same platform or infrastructure with data-intensive applications consuming high bandwidth and lot of Central Processing Unit (CPU) resources and memory. We can train the Node Manager using machine learning techniques to make scaling decisions for the virtual machine node (to scale up or down its resources) or for the entire cluster (add or remove one or more virtual machine nodes) based on the information Node Manager has aggregated for the virtual machine node or the cluster respectively. Different parameters in the information aggregated by the Node Manager will be assigned different weights in accordance with the Machine Learning algorithm used for training the Node Manager. Also other option is for Node Manager to make scaling decisions purely based on empirical data we derived from running the heavy load tests on the platform or infrastructure. The main purpose of running the heavy load tests on the platform is to derive the platform configuration for stable state of the virtual machine in the cluster given its current load and we can use predictive modelling to predict future loads for the virtual machine or the cluster and make scaling decisions based on that. The above novel technique of making scaling decisions for a virtual machine or a cluster of virtual machine nodes is the claim for this invention.

Description:
Metric Based Cluster Scaling

In this invention we have a group of virtual machine nodes in a cluster.

There is one virtual machine node called Node Manager which can be one of the virtual machine nodes in the cluster or outside of it. Node Manager periodically aggregates the information for each virtual machine node including Central Processing Unit (CPU) load and idle time of the virtual machine node, Memory usage of the virtual machine node, number of connections handled by the virtual machine node, etc. Now based on these parameters Node Manager makes decisions on whether to scale up or down the resources for a particular virtual machine or to add or remove one or more virtual machine nodes in the cluster. To make sure Node Manager takes appropriate decisions we run heavy load tests on the same platform or infrastructure with data-intensive applications consuming high bandwidth and lot of Central Processing Unit (CPU) resources and memory. We can train the Node Manager using machine learning

techniques to make scaling decisions for the virtual machine node (to scale up or down its resources) or for the entire cluster (add or remove one or more virtual machine nodes) based on the information Node Manager has aggregated for the virtual machine node or the cluster respectively. Different parameters in the information aggregated by the Node Manager will be assigned different weights in accordance with the Machine Learning algorithm used for training the Node Manager. Also other option is for Node Manager to make scaling decisions purely based on empirical data we derived from running the heavy load tests on the platform or infrastructure. The main purpose of running the heavy load tests on the platform is to derive the platform configuration for stable state of the virtual machine in the cluster given its current load and we can use predictive modelling to predict future loads for the virtual machine or the cluster and make scaling decisions based on that.