Fine-Grained Power Control for High Capacity Computer Clusters
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional data center designs face inefficiencies due to partially filled racks, leading to lower power utilization efficiency (PUE) and higher total cost of ownership (TCO) because of estimated peak power requirements from modern servers with multiple power-hungry modules, resulting in fixed costs for maintenance and operation despite unused capacity.
Innovation Solution
A system for fine-grained power management that uses a task distributor to optimize power consumption by placing inactive servers, drives, channels, and integrated circuits into power-saving modes based on a global power state, adjusting voltage, and shutting off power islands, thereby maintaining power budgets and improving Quality of Service (QoS).
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If servers are densely packed in racks to improve space utilization, then rack capacity increases, but power consumption exceeds the power budget
Solution Approach 1:
The patent segments power management to the integrated circuit level within drives, channels, and storage devices. Individual power switches control power delivery to specific ICs, allowing selective powering down of inactive components while keeping active ones running, thus enabling higher server density without exceeding power budgets.
Solution Approach 2:
The system implements fine-grained power control where different power states are applied to different components based on their activity. Power switches individually control power to inactive ICs, channels, or drives while maintaining power to active components, optimizing local power consumption to support increased overall density.
2Loss of energy
If estimated peak power is used to determine rack capacity, then power budget is preserved, but rack space utilization decreases
Solution Approach 1:
The system transitions from static power budgeting based on peak estimates to dynamic power management. Power states are adjusted in real-time based on actual workload and activity levels of components, allowing the rack to accommodate more servers by dynamically controlling power delivery to match actual needs rather than worst-case estimates.
Solution Approach 2:
The patent changes the power delivery parameter at the integrated circuit level through controlled impedance power switches. By adjusting power state parameters (on/off, active/standby) for individual ICs based on activity detection, the system optimizes power budget utilization to support higher rack density.
3Use of energy by moving object
If fixed rack footprint is maintained with partial filling, then power consumption is reduced, but operational efficiency and TCO increase
Solution Approach 1:
The system implements feedback mechanisms where activity detection circuits monitor the state of storage devices and channels, and this information feeds back to power control logic. This feedback enables automatic adjustment of power states to match actual workload, ensuring operational efficiency is maintained while optimizing power consumption and rack utilization.
Solution Approach 2:
Components equipped with activity detection circuits can autonomously determine their own power state needs. Inactive drives, channels, or ICs self-regulate their power consumption by triggering power switches to enter low-power modes, eliminating the need for external power management intervention and improving overall operational efficiency.
Data Source
AI summary
One embodiment of the present invention provides a system for a fine-grained power management. The system receives, by a first server, a task assigned based on a global power state, wherein the first server includes one or more drives, a drive includes one or more channels, and a channel includes one or more integrated circuits. The system places an inactive drive into a power-saving mode. The system places an inactive channel of an active drive into the power-saving mode by using a power switch associated with each integrated circuit in the inactive channel. The system places an inactive integrated circuit of an active channel into the power-saving mode by using a power switch associated with the inactive integrated circuit. The system updates a data structure storing the global power state based on a current power state of the first server, the drives, the channels, and the integrated circuits.