Neural Network Workload Prediction in Hyper-Converged Infrastructure
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In hyper-converged infrastructure, administrators face challenges in accurately predicting workload usage demands, leading to potential over-provisioning or under-provisioning of resources, which results in unnecessary costs, performance issues, and potential service level agreement violations.
Innovation Solution
A neural network-based usage prediction system, specifically utilizing a Long Short-Term Memory Recurrent Neural Network (LSTM-RNN), generates predictions for CPU, memory, and storage resource allocation, which are validated and translated into resource adjustments to optimize resource allocation before demand peaks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If historical usage data is used to predict future demand, then resource allocation can be made, but the prediction accuracy deteriorates due to unforeseen events and changing patterns
Solution Approach 1:
The patent transforms the prediction approach by changing from using only historical usage parameters to incorporating external event parameters. The system collects both historical workload usage data and external event data (such as holidays, promotions, or market changes), then feeds these combined parameters into machine learning models to generate more accurate future usage predictions. This parameter expansion resolves the contradiction by enabling predictions to adapt to unforeseen events while maintaining allocation efficiency.
Solution Approach 2:
The patent introduces external event data as an intermediary factor between historical usage patterns and future demand predictions. Instead of directly extrapolating from historical usage alone, the system uses external events as mediating variables that explain deviations from historical patterns. This intermediary approach allows the system to account for unforeseen events and changing business conditions, thereby improving prediction accuracy without sacrificing allocation efficiency.
2Reliability
If physical resources are over-allocated to workloads, then system reliability is improved, but cost increases due to unnecessary resource provisioning
Solution Approach 1:
The patent applies preliminary action by using machine learning models to predict future usage demands before they actually occur. The system proactively adjusts resource allocation in advance based on predicted usage patterns, allowing the system to be prepared for upcoming demand spikes without over-provisioning for unexpected events. This preliminary predictive adjustment resolves the contradiction by enabling reliable resource allocation that matches actual future demand rather than relying on conservative historical over-allocation.
Solution Approach 2:
The patent implements feedback mechanisms where actual usage data and external event information continuously refine the machine learning models. The system learns from actual outcomes and adjusts its prediction algorithms to better forecast future usage patterns. This feedback loop enables progressively more accurate predictions, allowing the system to reduce over-allocation while maintaining reliability through data-driven optimization rather than static conservative provisioning.
3Quantity of substance
If physical resources are under-allocated to workloads, then cost is reduced, but system performance deteriorates due to insufficient resources
Solution Approach 1:
The patent applies dynamics by transitioning from static resource allocation based on historical averages to dynamic allocation based on predicted future usage patterns. The system continuously updates resource allocation decisions based on real-time external event data and evolving usage patterns captured by machine learning models. This dynamic approach allows the system to optimize resource allocation in real-time, preventing both over-provisioning and under-provisioning by adapting to changing conditions while maintaining performance.
Data Source
AI summary
Systems and methods for analyzing the usage of a set of workloads in a hyper-converged infrastructure are disclosed. A neural network model is trained based upon historical usage data of the set of workloads. The neural network model can make usage predictions of future demands on the set of workloads to minimize over-allocation or under-allocation of resources to the workloads.