Energy-Aware Power-On Control for Computing Nodes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High-performance computing systems face inefficiencies in power management, leading to unnecessary energy consumption due to idle computing nodes waiting for boot or job completion, as accurate prediction of job execution time is challenging, making it difficult to optimize power-on timing and reduce energy waste.
Innovation Solution
An information processing device calculates energy consumption during boot latency and waiting times for computing nodes and performs power-on control when the waiting energy consumption equals or exceeds a threshold, optimizing power management by minimizing idle energy usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If free computing nodes are powered off to reduce energy consumption, then energy loss is reduced, but boot waiting time increases causing additional power loss
Solution Approach 1:
The job scheduler performs preliminary actions by predicting job end times and pre-powering on computing nodes before they are actually needed. The system calculates predicted end times based on scheduled usage lengths and uses this information to proactively manage power states, avoiding both unnecessary power consumption and boot waiting delays.
Solution Approach 2:
The system implements feedback mechanisms where the job scheduler continuously monitors actual job execution times and compares them against predicted end times. This feedback loop allows the system to refine its power management strategies, adjusting power-on timing based on actual performance data to minimize both energy consumption and waiting time.
2Reliability
If computing nodes wait for boot completion before executing jobs, then system reliability is improved, but energy consumption increases during waiting period
Solution Approach 1:
The system performs preliminary power-on actions for computing nodes based on predicted job end times. By pre-powering on nodes before they are actually needed, the system ensures immediate availability when jobs complete, eliminating waiting periods while maintaining reliability.
Solution Approach 2:
The power management system dynamically adjusts computing node states based on real-time job execution status and predicted completion times. The system transitions nodes between powered-off and powered-on states optimally, creating a dynamic power management approach that adapts to actual system conditions rather than following fixed schedules.
3Productivity
If power-on timing is advanced to reduce waiting time, then productivity is improved, but energy consumption increases due to premature booting
Solution Approach 1:
The system performs preliminary power-on actions based on predicted end times rather than actual end times. This allows the system to prepare computing nodes in advance without causing premature booting, as the predictions are adjusted to account for actual job completion timing, thereby avoiding unnecessary energy consumption.
Solution Approach 2:
The system changes the parameter used for power-on timing from fixed scheduled end times to predicted actual end times based on scheduled usage lengths. This parameter adjustment allows optimal power management by aligning power-on timing with actual job completion expectations, improving productivity while minimizing energy waste.
Data Source
AI summary
An information processing device includes a processor that calculates a first energy consumption of a first computer during a boot latency when the first computer has already booted and is not executing any job. The boot latency is time taken for boot of a second computer scheduled to execute a second job with the first computer after a first job executed by the first computer. The processor calculates a second energy consumption of the second computer during a waiting time when the second computer has already booted and is not executing any job. The waiting time is obtained by subtracting the boot latency from a time difference between a scheduled end time of the first job and a present time. The processor powers on the second computer when power of the second computer is off and when the second energy consumption becomes equal to or less than a threshold value.