Throughput-Optimized Power Capping for Co-Located QoS Workloads

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing power capping systems struggle to manage throughput-oriented and latency-sensitive workloads co-located in data centers, as they require fine-grained mechanisms that are not amenable to performance throttling, posing challenges in maintaining power safety and meeting service-level objectives.

Innovation Solution

A system that employs reactive and proactive capping mechanisms, utilizing CPU bandwidth control and load shaping policies to throttle throughput-oriented workloads while exempting latency-sensitive workloads, ensuring power safety and minimizing performance impact, with a two-threshold scheme and failover mechanisms.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of energy

If power capping actions are applied to throughput-oriented workloads, then power consumption is reduced, but workload completion deadlines may be missed resulting in lost revenue and diminished quality

Engineering Contradiction:
Improvepower consumptionVSAvoidworkload completion reliability
Core Design Contradiction:
Loss of energyVSReliability

Solution Approach 1:

The system applies differentiated power capping strategies to different workload types. Throughput-oriented workloads receive power capping actions while latency-sensitive workloads are exempted, creating local quality variations in power management policy across different workload categories co-located on the same server.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent segments workloads into distinct categories (throughput-oriented vs. latency-sensitive) and applies separate power management policies to each segment. This segmentation enables selective power capping on throughput workloads without affecting the reliability of latency-sensitive workloads that require guaranteed completion.

Inventive Principle:
Principle #1Segmentation

2Loss of energy

If power capping mechanisms are applied to latency-sensitive workloads, then power consumption is reduced, but service-level objectives are violated resulting in bad user experience

Engineering Contradiction:
Improvepower consumptionVSAvoidservice-level objective compliance
Core Design Contradiction:
Loss of energyVSEase of operation

Solution Approach 1:

The system implements local quality by creating distinct power management zones: latency-sensitive workloads receive premium treatment with power capping exemptions to maintain SLO compliance, while throughput-oriented workloads receive standard power capping treatment. This local differentiation resolves the contradiction by protecting SLO compliance for critical workloads.

Inventive Principle:
Principle #3Local quality

3Measurement precision

If fine-grained power capping mechanisms are implemented, then power control precision is improved, but system complexity increases

Engineering Contradiction:
Improvepower control precisionVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent employs dynamic power capping mechanisms that adjust power allocation in real-time based on workload characteristics and power availability. The system dynamically identifies workload types and applies appropriate capping strategies, achieving fine-grained power control precision through dynamic policy adjustment rather than static complex configurations.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The system implements self-service power management where the power capping mechanism automatically identifies workload types and applies appropriate policies without manual intervention. The workload classification and policy application occur autonomously, reducing operational complexity while maintaining precise power control.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP4407457B1A throughput-optimized, quality-of-service aware power capping system
Publication Date: 2025.07.09 GOOGLE LLC
  • EP4407457B1 patent drawingFigure 1A
  • EP4407457B1 patent drawingFigure 1B
  • EP4407457B1 patent drawingFigure 2A

AI summary

This disclosure describes a method to minimize disruption for throughput oriented jobs in power oversubscription services with a dynamic control. The mechanism controls power in a hardware-agnostic way, and the policy employs a multi-threshold approach that balances power safety with workload impact. Moreover, an alternative control mechanism ensures proper system operation while power measurements are unavailable.