Multivariate time series machine learning model for classifying network-intensive cloud workloads
Techniques are described for a compute optimizer service of a cloud provider network that uses a multivariate time series machine learning (ML) model to identify network-intensive cloud workloads and to provide recommendations for optimizing such workloads. A network-intensive workload, for example, broadly represents any computing workload that uses relatively more network resources than typical workloads (e.g., such as load balancers, web servers, and the like). The ML model combines a set of weak learners and a meta-model on top of the weak learners to infer the final predictions of whether workloads are network-intensive workloads. The ML model identifies network-intensive workloads in part by recognizing the patterns within the time series data reflecting certain resource usage patterns rather than only deriving embedded statistical information, thereby enabling the model to more accurately identify network-intensive workloads.
1 . A computer-implemented method comprising:
receiving, by a compute optimizer service of a cloud provider network, a request to obtain optimization recommendations related to one or more workloads deployed by a user to the cloud provider network, wherein the one or more workloads execute using computing resources provided by the cloud provider network;
obtaining, by the compute optimizer service, time series data reflecting performance metrics of a workload of the one or more workloads over time, the time series data including multiple different dimensions of performance metrics data;
utilizing a multi-level classification process to generate a probability value indicative of whether the workload is network-intensive, comprising:
executing multiple feature extraction processes to yield a plurality of sets of features based on the time series data, wherein at least one set of features is generated based on applying multiple different convolutional kernels to the time series data;
providing the sets of features to an ensemble of heterogeneous classifiers that were independently trained using different feature extraction processes to obtain a plurality of preliminary result values, wherein each preliminary result value indicates a first probability that the workload is a network-intensive workload;
providing the plurality of preliminary result values to a meta-classifier configured to combine the preliminary result values to obtain a final result value, wherein the final result value comprises a second probability that the workload is a network-intensive workload based on the plurality of preliminary result values;
determining, based on the second probability exceeding a threshold, that the workload is a network-intensive workload;
generating, based on the second probability, a recommendation to optimize at least one computing resource used to execute the workload, the recommendation including a recommendation to use a different type of virtual machine instance for the workload; and
causing display of the recommendation.
2 . The computer-implemented method of claim 1 , wherein the time series data includes data points reflecting, over a period of time for the workload, at least two of: central processing unit (CPU) utilization, inbound network throughput, outbound network throughput, inbound network traffic size, outbound network traffic size, inbound network packet count, network latency, packet loss, or outbound network packet count.
3 . The computer-implemented method of claim 1 , wherein the applying of multiple different convolutional kernels to the time series data comprises configuring each convolutional kernel with a set of parameters, wherein the parameters are either randomly generated from fixed distributions or selected from a predefined set of values, and wherein one or more features are generated from the output of each convolutional kernel.
4 . The computer-implemented method of claim 3 , wherein the one or more features generated from the output of each convolutional kernel comprise at least one of: a maximum value of the output, a proportion of positive values in the output, or a statistical or positional characteristic derived from the output of the convolutional kernel.
5 . A computer-implemented method comprising:
obtaining, by one or more computing devices, time series data corresponding to performance metrics of a workload deployed by a user to a cloud provider network, wherein the workload executes using at least one virtual machine (VM) instance provided by the cloud provider network;
generating a plurality of sets of features based on the time series data, wherein at least one set of features is generated based on applying multiple different convolutional kernels to the time series data;
providing the plurality of sets of features based on the time series data to an ensemble of heterogeneous classifiers that were independently trained using different feature extraction processes to obtain a plurality of preliminary result values, wherein each preliminary result value indicates a first probability that the workload is a network-intensive workload;
providing the plurality of preliminary result values to a meta-classifier configured to combine the preliminary result values to obtain a final result value, wherein the final result value indicates a second probability that the workload is a network-intensive workload based on the plurality of preliminary result values;
determining, based on the second probability, that the workload is a network-intensive workload;
identifying, based on the determination that the workload is a network-intensive workload, a different type of VM instance that is more suitable for the workload; and
causing display of a user interface providing a recommendation identifying the different type of VM instance.
6 . The computer-implemented method of claim 5 , wherein the time series data includes data points reflecting, over a period of time for the workload, at least two of: central processing unit (CPU) utilization, inbound network throughput, outbound network throughput, inbound network traffic size, outbound network traffic size, inbound network packet count, network latency, packet loss, or outbound network packet count.
7 . The computer-implemented method of claim 5 , wherein the recommendation comprises a
a recommendation to migrate the workload from the VM instance to the different type of VM instance, wherein the different type of VM instance is a network-optimized VM instance type.
8 . The computer-implemented method of claim 5 , wherein the workload is one of a plurality of workloads deployed by the user to the cloud provider network, and wherein the method further comprises:
determining whether each workload of the plurality of workloads is a network-intensive workload; and
causing display of an indication of workloads from the plurality of workloads identified as being network-intensive workloads.
9 . The computer-implemented method of claim 5 , wherein the meta-classifier is a k-nearest neighbor (KNN) classifier, a random forest classifier, or a decision tree classifier.
10 . The computer-implemented method of claim 5 , wherein the ensemble of classifiers includes at least two different classifiers from: a random forest classifier, a naive Bayes classifier, a ridge classifier, a decision tree classifier, a logistic regression-based classifier, a neural network classifier, a k-nearest neighbor (KNN) classifier, or a Diverse Representation Canonical Interval Forest Classifier (DrCIF).
11 . The computer-implemented method of claim 5 , wherein the time series data is collected by a monitoring service of the cloud provider network, wherein the monitoring service collects the time series data by monitoring execution of the workload over time.
12 . The computer-implemented method of claim 5 , wherein the ensemble of classifiers is trained based on one or more training sets reflecting performance metrics of example network-intensive workloads, wherein the example network-intensive workloads include at least one of: a load balancer, or a web server.
13 . The computer-implemented method of claim 5 , further comprising:
receiving an input requesting to implement the recommendation;
launching a new virtual machine instance of the different type; and
migrating the workload to the new virtual machine instance.
14 . The computer-implemented method of claim 5 , wherein the applying of multiple different convolutional kernels to the time series data comprises configuring each convolutional kernel with a set of parameters, wherein the parameters are either randomly generated from fixed distributions or selected from a predefined set of values, and wherein one or more features are generated from the output of each convolutional kernel.
15 . The computer-implemented method of claim 14 , wherein the one or more features generated from the output of each convolutional kernel comprise at least one of: a maximum value of the output, a proportion of positive values in the output, or a statistical or positional characteristic derived from the output of the convolutional kernel.
16 . A system comprising:
a first one or more electronic devices to implement a compute optimizer service in a multi-tenant provider network, wherein the compute optimizer service includes instructions that upon execution cause the compute optimizer service to:
obtain, from a monitoring service of the multi-tenant provider network, time series data corresponding to performance metrics of a workload deployed by a user to a cloud provider network, wherein the workload executes using at least one virtual machine (VM) instance provided by the cloud provider network,
generate a plurality of sets of features based on the time series data, wherein at least one set of features is generated based on applying multiple different convolutional kernels to the time series data;
provide the plurality of sets of features based on the time series data to an ensemble of heterogeneous classifiers trained using different feature extraction processes to obtain a plurality of preliminary result values, wherein each preliminary result value indicates a first probability that the workload is a network-intensive workload,
provide the plurality of preliminary result values to a meta-classifier configured to combine the preliminary result values to obtain a final result value, wherein the final result value indicates a second probability that the workload is a network-intensive workload based on the plurality of preliminary result values,
determine, based on the second probability, that the workload is a network-intensive workload,
identify, based on the determination that the workload is a network-intensive workload, a different type of VM instance that is more suitable for the workload, and
cause display of a user interface providing a recommendation identifying the different type of VM instance; and
a second one or more electronic devices to implement the monitoring service in the multi-tenant provider network, wherein the monitoring service includes instructions that upon execution cause the monitoring service to:
monitor execution of the workload, and
collect the time series data corresponding to performance metrics of the workload.
17 . The system of claim 16 , wherein the time series data includes data points reflecting, over a period of time for the workload, at least two of: central processing unit (CPU) utilization, inbound network throughput, outbound network throughput, inbound network traffic size, outbound network traffic size, inbound network packet count, network latency, packet loss, or outbound network packet count.
18 . The system of claim 16 , wherein the recommendation comprises
a recommendation to migrate the workload from the VM instance to the different type of VM instance, wherein the different type of VM instance is a network-optimized VM instance type.
19 . The system of claim 16 , wherein the workload is one of a plurality of workloads deployed by the user to the cloud provider network, and wherein the compute optimizer service further includes instructions that upon execution cause the compute optimizer service to:
determine whether each workload of the plurality of workloads is a network-intensive workload; and
cause display of an indication of workloads from the plurality of workloads identified as being network-intensive workloads.
20 . The system of claim 16 , wherein the compute optimizer service, as part of the generation of the plurality of sets of features, is to apply multiple different convolutional kernels to the time series data, wherein each convolutional kernel is configured with a set of parameters, the parameters being either randomly generated from fixed distributions or selected from a predefined set of values, and wherein one or more features are generated from the output of each convolutional kernel.