Predictive SDDC Scaling Using ML Forecasting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current software-defined data center (SDDC) scaling processes are inefficient due to reactive decision-making, lack of cost-benefit analysis, and failure to account for input/output (IO) utilization, leading to suboptimal resource allocation and potential performance issues.
Innovation Solution
Implement a predictive scaling method that stores resource utilization information, uses machine-learning models to forecast future utilization rates, and determines the required time for scaling operations, comparing the predicted benefit period to a customer-set threshold to optimize resource allocation and reduce costs or ensure performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If reactive scaling based on moving average is used, then system simplicity is maintained, but scaling response time lags behind actual demand causing poor performance
Solution Approach 1:
The patent applies preliminary action by predicting future resource utilization rates before scaling decisions are made. The machine learning model forecasts upcoming demand spikes, allowing the system to provision resources in advance of actual demand, eliminating the lag inherent in reactive moving average approaches.
Solution Approach 2:
The patent implements feedback by continuously monitoring actual resource utilization and comparing it against predicted rates. This feedback loop allows the system to refine its predictions and adjust scaling decisions dynamically, improving response time while maintaining manageable complexity through iterative optimization.
2Loss of energy
If host removal process is initiated, then resource costs are reduced, but performance deteriorates if demand rises during the removal process
Solution Approach 1:
The patent applies preliminary action by predicting future demand patterns before initiating host removal. The machine learning model forecasts whether demand will rise during the removal process, allowing the system to delay or cancel removal operations when predicted demand increases are detected, thus preventing performance deterioration while still achieving cost optimization when appropriate.
3Device complexity
If storage space utilization is used as the only scaling metric, then monitoring simplicity is maintained, but IO performance issues are not detected leading to suboptimal scaling decisions
Solution Approach 1:
The patent applies segmentation by separating storage metrics into distinct components: storage space utilization, IO utilization, and IO bandwidth. This segmentation allows the system to monitor multiple dimensions of storage performance independently, providing comprehensive visibility into IO performance issues while maintaining manageable monitoring complexity through modular metric collection and analysis.
Data Source
AI summary
Examples described herein include systems and methods for efficiently scaling an SDDC. An example method can include storing resource utilization information for a variety of resources of the SDDC. The example method can also include predicting a future resource utilization rate for the resources and determining that a predicted utilization rate is outside of a desired range. The system can determine how long it would take to perform the scaling, including adding or removing a host and performing related functions such as load balancing or data transfers. The system can also determine how long the scaling is predicted to benefit the SDDC to ensure that the benefit is sufficient to undergo the scaling operation. If the expected benefit is greater than the benefit threshold, the system can perform the scaling operation.