Data Center Power Forecasting for Hotspot-Aware Workload Scheduling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional approaches lack the ability to dynamically predict power requirements of servers based on historical usage patterns, leading to inefficient workload scheduling and increased costs in data centers due to overloading or under-utilization of power supplies.
Innovation Solution
Utilizing machine learning models, such as FBProphet, SARIMA, Holt-Winters-ES, and GRU Networks, to forecast power consumption and identify hotspots (over-utilization) and coldspots (under-utilization) in data centers by calculating exponential moving averages, enabling efficient workload scheduling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a one size fits all power supply strategy is used, then infrastructure planning is simplified, but cost increases and performance/health of the data center deteriorates
Solution Approach 1:
The patent implements dynamic power supply management by continuously monitoring server power consumption patterns and adjusting power allocation in real-time. The system transitions from static one-size-fits-all power supplies to dynamic power adjustment based on actual workload demands, optimizing energy efficiency while maintaining infrastructure manageability through automated control mechanisms.
Solution Approach 2:
The system changes power supply parameters dynamically based on predicted power consumption patterns. By analyzing historical data and using machine learning models, the system adjusts power allocation parameters for each server based on its specific workload characteristics, replacing the fixed parameter approach with adaptive parameter adjustment to reduce energy waste.
2Reliability
If power supplies are overloaded to meet peak demands, then service availability is maintained, but energy waste and operational costs increase
Solution Approach 1:
The system performs preliminary power consumption prediction using machine learning models trained on historical data. By forecasting future power demands before they occur, the system can proactively adjust power allocation to match actual needs, preventing both overloading and under-utilization scenarios, thereby reducing energy waste while maintaining service availability.
Solution Approach 2:
The patent implements feedback mechanisms that continuously monitor actual server power consumption and compare it against predictions. This feedback loop enables the system to learn from actual usage patterns and refine its predictions, allowing for more accurate power allocation that prevents energy waste while ensuring service reliability through adaptive control.
3Loss of energy
If power supplies are under-utilized to reduce costs, then energy waste is reduced, but service performance and reliability deteriorate
Solution Approach 1:
The system dynamically adjusts power supply allocation based on real-time and predicted workload demands. Rather than maintaining fixed under-utilized power supplies, the system adapts power capacity to match actual server needs, ensuring sufficient power is available when required while avoiding waste during low-demand periods through flexible, demand-driven power management.
4Loss of energy
If dynamic power prediction based on historical patterns is implemented, then power usage optimization is achieved, but system complexity and computational requirements increase
Solution Approach 1:
The system implements self-service power management where servers automatically report their own power consumption patterns and the system autonomously generates predictions and adjustments without requiring manual intervention. This automation reduces the operational complexity burden despite the increased computational complexity of the prediction models, as the system manages itself rather than requiring complex manual control processes.
Data Source
AI summary
Systems and methods are provided for using historic input power periodic data from a server in an IT data center to train a machine learning (ML) model to obtain forecasted power consumption data of the server for a future time period. Time windows of hotspots or coldspots are then identified in the forecasted power consumption data, hotspots being defined as areas or regions of over-utilization in a time series data, and coldspots being defined as areas or regions of under-utilization in a time series data. The hotspots and coldspots are identified by calculating an exponential mean average (EMA) of the forecasted power consumption data, taking points above the EMA as hotspots and points below the EMA as coldspots. The identified hotspots and coldspots can be used to schedule workloads for a server or a data center, to more efficiently plan existing workloads, or to introduce new workloads at more optimal time periods.