Distributed Job Power Estimation Using Node Calibration Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current distributed computer systems inaccurately estimate power needs, leading to delayed job starts and inefficient resource allocation due to overestimation of power requirements, which affects performance and causes potential power overconsumption.
Innovation Solution
A method to estimate power performance for jobs in distributed computer systems based on calibration data, including workload type, node selection, and operational frequency, providing accurate power allocation and management within limited power budgets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If power capping is applied to adhere to limited power and energy budgets, then power consumption is controlled, but job performance is negatively impacted due to inaccurate power estimation
Solution Approach 1:
The system implements feedback by continuously monitoring actual power consumption of nodes and comparing it with estimated power requirements. This feedback loop enables dynamic adjustment of power allocation and job scheduling decisions, allowing the system to maintain accurate power estimation while adhering to power budgets without unnecessarily delaying job starts.
Solution Approach 2:
The patent changes the estimation parameter from static TDP values to dynamic actual power consumption measurements. By collecting and analyzing real power consumption data from nodes under various workloads, the system adjusts power estimation parameters to reflect actual usage patterns, thereby improving both power control accuracy and job performance.
2Device complexity
If TDP values are used to estimate power needs, then power estimation is simplified, but job start is delayed due to overestimation of power requirements
Solution Approach 1:
The system performs preliminary action by pre-collecting actual power consumption data from nodes during calibration phases before jobs are scheduled. This advance data collection creates a database of real power requirements that can be quickly referenced during job scheduling, avoiding both overestimation delays and complex real-time measurements when jobs need to start.
Solution Approach 2:
The patent uses cheap, easily obtainable TDP specifications as initial estimates but replaces them with more accurate actual power consumption measurements whenever available. This approach allows the system to use simple data sources when necessary while transitioning to more accurate but resource-intensive measurement methods when accuracy is critical.
3Ease of operation
If TDP is used as the basis for power allocation, then power distribution is straightforward, but power overconsumption may occur when actual power needs exceed TDP estimates
Solution Approach 1:
The system substitutes the mechanical/specification-based TDP estimation approach with an information-based approach using actual power consumption measurements. By replacing static manufacturer specifications with dynamic real-world data, the system maintains ease of operation through automated data collection while dramatically improving power allocation accuracy and reliability.
Solution Approach 2:
The patent creates a composite estimation model that combines multiple data sources: TDP specifications, actual power consumption measurements under various workloads, and node configuration information. This composite approach leverages the simplicity of TDP while incorporating the accuracy of real measurements, resulting in a robust power allocation system that is both easy to operate and highly reliable.
4Productivity
If over-estimation of startup power is avoided through accurate estimation, then job start delay is reduced, but more complex power monitoring and calibration systems are required
Solution Approach 1:
The system implements self-service by having nodes automatically report their own power consumption data to the management system. Nodes perform self-calibration by monitoring their own power usage under various workloads and storing this information locally and remotely. This self-service approach reduces the need for complex external monitoring infrastructure while enabling accurate power estimation that improves job execution efficiency.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A non-transitory computer readable storage medium having stored thereon instructions executable by one or more processors to perform operations including: receiving a plurality of input parameters including (i) a workload type, (ii) a list of selected nodes belonging to a distributed computer system, and (iii) a list of frequencies; responsive to receiving the plurality of workload parameters, retrieving calibration data from a calibration database; generating a power estimate based on the plurality of workload parameters and the calibration data; and providing the power estimate to a resource manager is shown. Alternatively, the input parameters may include (i) a workload type, (ii) a list of selected nodes belonging to a distributed computer system, and (iii) an amount of available power, wherein the estimator may provide an estimation of the frequency at which the nodes should operate to utilize as much of the available power without exceeding the available power.