Parallel Processing Power Management via Job-Aware Node Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current parallel processing systems face challenges in reducing electric power consumption without compromising job completion rates, as stopping or underclocking nodes that execute high-power jobs can lead to incomplete tasks and subsequent re-execution, resulting in excessive power consumption.
Innovation Solution
An apparatus that stores power consumption data for nodes and jobs, identifies the job with the largest power consumption, and executes processes to reduce power consumption based on the remaining execution time, either by reducing the power of the node executing the job or a different node, to maintain efficient operation and completion rates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Use of energy by moving object
If nodes are stopped or underclocked to reduce power consumption, then power consumption is reduced, but job completion rate deteriorates
Solution Approach 1:
The system dynamically adjusts node operation states based on real-time power thresholds and job characteristics. Nodes transition between active, underclocked, and stopped states depending on current power consumption levels and job remaining execution times, allowing adaptive optimization without compromising job completion
Solution Approach 2:
The system identifies jobs with long remaining execution times in advance and proactively selects their nodes for power reduction. By predicting which jobs can tolerate node state changes without missing deadlines, the system prevents power consumption increases from re-execution while maintaining job completion rates
2Use of energy by moving object
If nodes are stopped or underclocked to reduce power consumption, then power consumption is reduced, but re-execution of incomplete jobs occurs leading to increased total power consumption
Solution Approach 1:
The system continuously monitors node power consumption, job progress, and remaining execution times. Based on this feedback, it dynamically determines whether to reduce node power states, ensuring that power reduction actions only occur when they will not cause job completion failures and subsequent re-execution
Solution Approach 2:
By identifying jobs with sufficient remaining execution time buffers before applying power reduction, the system prevents incomplete job executions. This preliminary assessment ensures that nodes are only underclocked or stopped when it is safe to do so, avoiding the energy waste of re-execution
Data Source
AI summary
A system includes a management device and nodes that execute plural jobs in parallel. When a total electric-power consumption of the nodes reaches a threshold, the management device extracts a first job of the largest electric-power consumption from among the plural jobs, based on information about an electric-power consumption of each node and information about the plural jobs that are executed in parallel by the nodes. The management device reduces the electric-power consumption of a first node that executes the first job when a remaining execution time of the first job, which indicates a period of time from a current time until a scheduled end time of the first job, is longer than or equal to a predetermined time, and reduces the electric-power consumption of a second node that does not execute the first job, when the remaining execution time of the first job is shorter than the predetermined time.