Datacenter Load Shedding with Workload-Aware Response Levels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional datacenter power management techniques lead to suboptimal load shedding, as operators lack insight into workloads and customer impacts during power reduction actions, resulting in inefficient power consumption adjustments and potential broader disruptions.
Innovation Solution
A method involving a computer system that identifies workloads on hosts, determines response levels for power reduction actions, and applies these actions dynamically to minimize impact on customers and workloads while maintaining power balance within the datacenter.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If manual load shedding is performed by shutting down host racks or individual devices one by one, then power consumption is reduced, but the operator lacks insight into workloads and customer impacts resulting in suboptimal power management
Solution Approach 1:
The system implements feedback by continuously monitoring workload information, customer impact data, and power consumption metrics. This feedback loop enables operators to see the direct consequences of load shedding actions on workloads and customers, allowing for informed decision-making about which devices to power down and what impacts to expect.
Solution Approach 2:
The system introduces an intermediary layer between the manual shutdown action and the actual power reduction. This intermediary collects and processes information about workloads and customer impacts, presenting this data to operators before they execute load shedding decisions, thereby bridging the information gap.
2Loss of energy
If emergency shut off switch is pressed to turn off power to the whole or large portion of the datacenter, then power consumption is rapidly reduced, but widespread power failure occurs causing significant disruption to downstream devices
Solution Approach 1:
The system segments the datacenter power infrastructure into hierarchical levels (upstream and downstream devices). Instead of applying emergency shutdown to the entire system at once, operators can selectively power down specific segments or devices while maintaining power supply to others, thereby reducing overall power consumption without causing widespread failure.
Solution Approach 2:
The system enables dynamic, granular control over power distribution rather than static all-or-nothing shutdown. Operators can progressively power down devices based on real-time conditions, workload priorities, and customer impact assessments, allowing flexible adaptation to power constraints while maintaining service reliability.
3Loss of energy
If conventional load shedding methods are used without workload awareness, then power consumption is reduced, but customer experience deteriorates due to lack of targeted power management
Solution Approach 1:
The system applies local quality by enabling differentiated power management for different devices and workloads based on their specific characteristics, customer priorities, and operational importance. Instead of uniform load shedding, the system allows targeted power reduction on less critical devices while preserving power to high-priority workloads, thereby maintaining customer experience.
4Reliability
If power failure occurs in one datacenter, then cascading power failure can trigger other devices within the same datacenter or other datacenters, but automated orchestration can prevent this by coordinating power reduction actions
Solution Approach 1:
The power orchestration system implements multi-functionality by simultaneously performing multiple tasks: monitoring power consumption across hierarchical levels, collecting workload and customer impact data, presenting load shedding options to operators, and executing coordinated power reduction actions. This universal system prevents cascading failures while managing power constraints across the entire datacenter infrastructure.
Data Source
AI summary
Disclosed techniques relate to orchestrating power consumption reductions across a number of hosts. A number of response levels may be utilized, each having an association to a corresponding set of reduction actions. The impact to customers, hosts, and/or workloads can be computed at run time based on current and/or predicted conditions and workloads, and a particular response level can be selected based on the computed impact. These techniques enable a sufficient, but least impactful response to be employed.