Energy Usage Estimation in Distributed Computing Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for estimating energy usage in distributed computing systems are inaccurate due to the complex and cascading nature of operations, leading to inefficient energy allocation and cost management.
Innovation Solution
The use of latency data from segments of operations performed on behalf of users to infer the weight of operations relative to others, combined with a machine learning model trained on tracing data, to generate accurate energy usage estimates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional energy estimation methods are used in distributed computing systems, then the system can operate with simple tracking mechanisms, but the energy usage attribution to individual users becomes inaccurate
Solution Approach 1:
The patent segments the computing system into multiple hierarchical levels (users, applications, services, operations, segments) and tracks energy consumption at each level. This segmentation allows accurate attribution of energy usage to individual users by breaking down the complex system into manageable, traceable components, directly resolving the contradiction between measurement precision and system complexity.
Solution Approach 2:
The patent introduces an intermediary energy estimation component that uses machine learning models to process tracing data and generate energy usage estimates. This intermediary layer translates complex system operations into accurate energy attributions without requiring direct modification of all system components, thereby improving measurement precision while controlling system complexity.
2Measurement precision
If detailed tracing data is collected from all operations to improve energy estimation accuracy, then energy usage attribution becomes precise, but the computing overhead and processing burden increase
Solution Approach 1:
The patent implements preliminary action by pre-processing and storing tracing data in a structured format during system operations. This allows the energy estimation component to quickly retrieve and analyze relevant data without causing processing bottlenecks, thereby maintaining both measurement precision and system throughput.
Solution Approach 2:
The patent applies partial action by selectively collecting and processing only the most relevant tracing data needed for energy estimation, rather than analyzing every single system operation in detail. This selective approach maintains accurate energy attribution while minimizing the computing overhead and preserving system productivity.
3Measurement precision
If machine learning models are trained on comprehensive tracing data to improve energy estimation, then user-specific energy usage patterns are captured accurately, but the training time and computational resources required increase
Solution Approach 1:
The patent applies local quality by training specialized machine learning models for different user profiles, application types, and service categories rather than using a single generic model. This allows each model to capture specific energy usage patterns with high accuracy while keeping individual model sizes and training requirements manageable, thus improving user-specific estimation accuracy without excessive training time.
Solution Approach 2:
The patent performs preliminary action by pre-training models during off-peak hours using historical tracing data, and periodically updating them with new data. This approach captures user-specific energy patterns accurately while minimizing the impact on system productivity and avoiding lengthy training processes during operational periods.
Data Source
AI summary
Tracing data including a plurality of traces for a plurality of operations performed by a distributed computing system on behalf of a plurality of users of a distributed computing system during a period of time is identified. Each trace having latencies for a plurality of segments of a corresponding operation. A set of overall latencies comprising an overall latency for each segment is determined. A set of user latencies including a latency for each segment is determined for each of the plurality of users. A set of energy usage estimates including an energy usage estimate for one or more of the plurality of users is generated, by a processing device, based on the set of overall latencies and the set of user latencies using a machine learning (ML) model.