System and method of monitoring and measuring cluster performance hosted by an IAAS provider by means of outlier detection

Patent metadata

Field	Value
Publication number	US-9514387-B2
Application number	US-201314144980-A
Country	US
Kind code	B2
Filing date	Dec 31, 2013
Priority date	Sep 17, 2013
Publication date	Dec 6, 2016
Grant date	Dec 6, 2016

Abstract

The present disclosure is directed to a system for monitoring and analyzing operation of a widely distributed service operated by an Infrastructure-as-a-Service (IaaS) tenant but deployed on a set of virtual resources controlled by an independent IaaS provider. The set of virtual resources can be organized into clusters in which resources are expected to behave similarly to each other. Virtual resources that do not behave similar to peer resources in the same cluster, i.e., outliers, may be indicative of problems that need to be addressed. The monitoring system can collect performance metric data from virtual resources, and compare the performance of each virtual resource in a cluster with the performance of every other virtual resource in the cluster to detect outliers. This comparison can involve correlation analysis, ANOVA analysis, or regression analysis.

First claim

What is claimed is:

1. A system comprising non-transitory memory with programmable instructions for detecting outlier virtual resources that perform differently from peer virtual resources within a cluster of virtual resources that are expected to perform similarly, the virtual resources being provided by an independent Infrastructure-as-a-Service (IaaS) provider to an IaaS tenant for operating a widely distributed service, wherein the widely distributed service may be at least one of geographically dispersed, part of different communication networks, and disjoint, wherein the IaaS provider is responsible for selection of resources, wherein an operational capacity of the resources may change substantially and rapidly, and wherein the IaaS tenant has no direct control over, and limited visibility into, the selection of resources, the system comprising:
a data gateway configured to collect substantially live system-level metrics related to the operation of the cluster of virtual resources; and
an analysis module configured to:
compare the live system-level metrics for each virtual resource in the cluster of virtual resources against live system-level metrics for other virtual resources in the cluster of virtual resources, wherein a cluster is a group of resources expected to behave similarly; and
determine that a given virtual resource is an outlier based on the comparison of the live system-metrics between each virtual resource against other virtual resources, wherein an outlier virtual resource is determined if a difference between the live system-level metrics for the given virtual resource and the live system-level metrics for other virtual resources in the cluster of virtual resources exceeds a predetermined threshold, and wherein the determination that the given virtual resource is an outlier includes performing correlation analysis on the given virtual resource, performing ANOVA analysis on the given virtual resource, or performing regression analysis on the given virtual resource;

provide, at a user interface, a notification to a user that the given virtual resource is an outlier amongst the given virtual resource's peer virtual resources within the cluster of virtual resources.

2. The system of claim 1 , wherein the predetermined threshold is based on at least one of a resource type of the given virtual resource, a resource role of the given virtual resource, the live system-level metric being compared, and feedback from the IaaS tenant in response to previously determined outliers.

3. The system of claim 1 , wherein the analysis module is configured to perform correlation analysis, including:
compute a correlation matrix for the system-level-metrics for all resources in the cluster of virtual resources;
sum, for each virtual resource in the cluster of virtual resources, the correlations of the system-level metrics for that resource against the system-level metrics of all other peer virtual resources in the cluster of virtual resources; and
determine that a given virtual resource is an outlier if the sum for the given virtual resource is more than one quartile away from a median for the sum for all the virtual resources in the cluster of virtual resources.

4. The system of claim 1 , wherein the analysis module is configured to:
collapse the system-level metrics for all resources in the cluster of virtual resources by removing time labels;
detect whether there is an outlier by determining whether the collapsed system-level metrics for at least one virtual resource exhibits a substantially different mean and variance than means and variances for other virtual resources by performing the ANOVA analysis;
if an outlier is detected, determine which virtual resource is the outlier using the Tukey HSD test.

5. The system of claim 1 , wherein the analysis module is configured to perform regression analysis, including:
regress, for each virtual resource in the cluster of virtual resources, the live performance-metric data for all other peer virtual resources in the cluster of virtual resources as independent variables onto the live performance-metric data for the virtual resource being analyzed as the dependent variable;
compute the goodness-of-fit for the regression for each virtual resource in the cluster of virtual resources; and
determine that a given virtual resource is an outlier if the goodness-of-fit for the regressions for all other virtual peer resources in the cluster satisfy a pre-defined threshold, but the goodness-of-fit for the regression of the given virtual resource does not satisfy the pre-defined threshold.

6. The system of claim 1 , wherein the data gateway is configured to collect the substantially live system-level metrics at a frequency that is high enough to capture an expected rate of change of the set of virtual resources.

7. The system of claim 6 , wherein the frequency is once every five minutes.

Assignees

• Google Inc

Inventors

• Eaton Patrick Randolph

Classifications

• H04L43/0817 Primary

  by checking functioning · CPC title

• G06F18/231

  Hierarchical techniques, i.e. dividing or merging pattern sets so as to obtain a dendrogram · CPC title

• G06F11/3447

  Performance evaluation by modeling · CPC title

• H04L41/069

  using logs of notifications; Post-processing of notifications · CPC title

• G06K9/6219 Primary

  Physics · mapped topic