Wide-and-Deep Capacity Prediction for Cloud Cluster Provisioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing predictive models for cloud resource allocation are inaccurate, leading to either resource under-provisioning, resulting in poor service performance, or over-provisioning, causing resource waste, due to the dynamic nature of workload demands in cloud computing.
Innovation Solution
A machine learning model combining wide and deep learning techniques is used to predict future resource usage by integrating discrete and time-series data, allowing for accurate capacity unit predictions, which inform real-time cluster allocation decisions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If existing predictive models are used for resource allocation, then resource provisioning can be automated, but prediction accuracy is low leading to either under-provisioning or over-provisioning
Solution Approach 1:
The patent segments the prediction task into multiple components: short-term predictions (next 1-7 days) and long-term predictions (next 8-30 days), using different modeling approaches for each time horizon. This segmentation allows each model to be optimized for its specific time scale, improving overall prediction accuracy while maintaining automation.
Solution Approach 2:
The system performs preliminary actions by continuously training models on historical data and generating predictions before resource allocation decisions are made. The predictions are prepared in advance and used to guide proactive resource provisioning, preventing both under-provisioning and over-provisioning before they occur.
2Productivity
If resource allocation is adjusted frequently to match dynamic workload, then service performance improves, but resource waste increases due to over-provisioning
Solution Approach 1:
The patent implements dynamic resource allocation by continuously updating predictions based on the latest workload data and adjusting resource provisioning accordingly. The system adapts to changing workload patterns in real-time, optimizing service performance while avoiding static over-provisioning that would waste resources during low-demand periods.
Solution Approach 2:
The system incorporates feedback mechanisms where actual resource usage and workload patterns are monitored and fed back into the prediction models. This feedback loop allows the system to learn from past allocations, refine predictions, and continuously improve resource allocation decisions to balance performance and waste reduction.
3Loss of energy
If resource allocation is conservative to avoid over-provisioning, then resource efficiency improves, but service performance degrades due to under-provisioning
Solution Approach 1:
The system takes preliminary action by generating predictions in advance that anticipate future resource needs. This allows for proactive resource allocation that ensures sufficient capacity is available before demand spikes occur, maintaining service reliability while avoiding the need for conservative over-provisioning.
Solution Approach 2:
The patent changes key parameters by using advanced machine learning models that capture complex workload patterns and dependencies. These parameter changes in the prediction approach enable more accurate estimation of true resource needs, allowing efficient allocation without compromising service performance through under-provisioning.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Real-time workload for an application is converted into a capacity unit value. A trained machine-learning model receives the capacity unit data as input and generates a prediction of capacity usage for the future. Based on the prediction, additional clusters may be allocated for the application. A dataset for use in predicting future capacity unit usage by an application may be classified into two categories. A first category comprises time series data. A second category comprises service types, target user types, and other attributes with discrete characteristics. Discrete attributes are incorporated into a wide section and time-series data is integrated into a deep section. A wide and deep model combines results from the wide section and the deep section to generate a prediction of capacity units used by the application in future time periods. In response, an allocation system allocates a corresponding number of clusters.