Dynamic Power Allocation for Pooled Accelerators
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Data center architectures face inefficiencies in power management due to fixed power supplies for pooled accelerators, leading to suboptimal load balancing and power allocation based on varying performance requirements of workloads, which increases total cost of ownership and hinders flexible power management.
Innovation Solution
A switch is configured to connect nodes to accelerators, managing power supply based on performance targets by monitoring current and threshold power levels, allowing dynamic power adjustment and redistribution among accelerators to optimize power usage and balance loads.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If fixed power supply is provided to pooled accelerators, then power management is simplified, but power allocation efficiency deteriorates and total cost of ownership increases
Solution Approach 1:
The patent implements dynamic power management by allowing power supply to accelerators to vary based on real-time performance requirements and workload demands. The system continuously monitors accelerator performance metrics and adjusts power allocation accordingly, transitioning from a static fixed power model to a dynamic adaptive power model that optimizes both efficiency and cost-effectiveness
Solution Approach 2:
The system changes the power parameter dynamically based on performance targets and workload characteristics. By adjusting power levels as a variable parameter rather than a fixed constant, the system achieves optimal power allocation efficiency while maintaining simplified management through automated control mechanisms
2Device complexity
If fixed power supply is provided to pooled accelerators, then power distribution is simplified, but load balancing capability deteriorates
Solution Approach 1:
The patent enables dynamic load balancing by continuously adjusting power distribution to accelerators based on real-time performance monitoring and workload analysis. The system adapts power allocation to match actual load requirements, improving load balancing capability while maintaining simplified distribution through automated dynamic control
Solution Approach 2:
The system implements feedback mechanisms that monitor accelerator performance and use this information to adjust power distribution dynamically. This closed-loop control enables the system to automatically balance loads across accelerators based on their actual performance characteristics and workload demands
3Device complexity
If power is allocated based on fixed supply, then power management is simpler, but energy efficiency deteriorates
Solution Approach 1:
The patent optimizes energy efficiency by changing power allocation from a fixed parameter to a dynamic parameter that adapts to actual workload requirements. The system adjusts power levels based on real-time performance metrics, ensuring that energy is allocated efficiently to match actual computational demands rather than providing constant fixed power
Data Source
AI summary
A computing device, a method and a system to control power. The computing device is configured to be used as part of a network fabric including a plurality of nodes and a plurality of pooled accelerators coupled to the nodes. The computing device includes: a memory storing instructions; and processing circuitry configured to perform the instructions. The processing circuitry is to receive respective requests from respective ones of the plurality of nodes, the requests addressed to a plurality of corresponding accelerators, each of the respective requests including information on a kernel to be executed by a corresponding accelerator, on the corresponding accelerator, and on a performance target for execution of the kernel. The processing circuitry is further to, based on the information in said each of the respective requests, control a power supply to the corresponding accelerator.