Job-Aware Power-Capping for Data Center Nodes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional power-capping methods in data centers, such as uniform node power-capping, do not consider application execution performance, leading to significant performance penalties that negate the benefits of power management.
Innovation Solution
Implementing a job-aware power-capping scheme where compute nodes are partitioned into power domains, with a global power dispatcher allocating power caps based on a total system budget, and runtime power-aware agents determining optimal node processing frequencies to maximize instructions per unit of power, adjusting power caps accordingly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by stationary object
If uniform node power-capping is implemented, then power consumption is reduced, but application execution performance deteriorates significantly
Solution Approach 1:
The patent implements non-uniform power-capping where different compute nodes receive different power caps based on their individual characteristics and workload requirements. The system determines application performance footprints for different nodes and allocates power budgets accordingly, allowing nodes that can maintain performance at lower power to do so, while providing higher power to nodes that require it for optimal performance.
Solution Approach 2:
The system dynamically adjusts power caps based on real-time monitoring of application performance footprints and workload characteristics. The power management is not static but adapts to changing conditions, allowing the system to optimize the balance between power consumption and performance as workloads evolve.
2Loss of energy
If power caps are reduced to manage power burden, then energy efficiency improves, but system productivity decreases
Solution Approach 1:
The system changes the power consumption parameters of compute nodes based on their measured performance footprints. By establishing performance-to-power relationships for different applications and nodes, the system can set appropriate power caps that maintain acceptable performance levels while reducing overall power consumption compared to uniform maximum power allocation.
Solution Approach 2:
The system continuously monitors application performance and uses this feedback to adjust power cap allocations. The performance footprint measurements provide feedback that informs power budget decisions, creating a closed-loop system that adapts to actual performance requirements rather than using fixed or estimated power allocations.
3Device complexity
If simple power-capping mechanisms are used, then power management complexity is reduced, but performance penalty increases
Solution Approach 1:
The system performs preliminary measurements of application performance footprints before finalizing power cap allocations. By characterizing the performance characteristics of different applications and their sensitivity to power changes in advance, the system can make more informed power management decisions that minimize performance penalties.
Solution Approach 2:
The patent introduces an intermediary power management system that sits between the hardware and applications, translating performance requirements into power cap decisions. This intermediary layer uses performance footprint data to mediate between power constraints and application needs, reducing the direct impact of power-capping on performance.
Data Source
AI summary
A technique includes an agent executing on a plurality of nodes while a job is being concurrently executed by the plurality of nodes. The plurality of nodes is power-capped by an existing node power consumption budget. The technique includes managing power consumption of the plurality of nodes. The managing includes the agent determining a performance footprint that is associated with execution of the job; and the managing includes the agent determining a second node power consumption budget based on the performance footprint. The second node power consumption budget is different than the existing node power consumption budget. The managing includes the agent providing a power consumption request to a global power dispatcher to set a new node power consumption budget for the plurality of nodes.