Adapting Pre-Trained Predictive Models for Distributed Resource Scheduling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Legacy techniques for implementing resource performance predictive models in hyperconverged distributed systems suffer from significant prediction errors during the initial learning phase, particularly when the model is trained in a different environment from the target system, leading to inefficient resource scheduling.
Innovation Solution
The use of pre-trained resource performance predictive models that are dynamically adapted to the target hyperconverged computing environment by modifying training model parameters based on the target system configuration and workload schedule, reducing prediction errors and improving scheduling efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a pre-trained resource performance predictive model is deployed to a target system with different configuration and workload characteristics, then the model can be quickly deployed without lengthy training, but the model exhibits large prediction errors during the initial learning phase
Solution Approach 1:
The model is pre-trained on a source system before deployment to the target system. This preliminary training action prepares the model with general resource performance patterns, enabling quick deployment while maintaining reasonable predictive capability. The pre-training occurs in advance, allowing the model to be ready for immediate use on the target system without requiring lengthy on-site training.
Solution Approach 2:
The patent modifies model parameters during deployment to adapt to the target system's specific configuration and workload characteristics. By changing parameters such as scaling factors, offsets, and other adjustable model variables, the system tailors the pre-trained model to match the target environment, thereby reducing prediction errors while maintaining quick deployment.
2Loss of time
If the model is trained in a computing environment different from the target system environment, then deployment time is reduced, but the training period becomes longer and prediction errors increase
Solution Approach 1:
Training is performed in advance on a source system that may have different characteristics from the target system. This preliminary training action allows the model to learn general resource performance patterns before deployment, reducing on-site training time while maintaining adequate predictive reliability through subsequent parameter adaptation.
Solution Approach 2:
Model parameters are modified during the deployment process to account for differences between the source training environment and the target system environment. This parameter adjustment compensates for environmental differences, maintaining prediction reliability despite training in a different computing environment.
3Ease of manufacture
If legacy techniques are used to deploy untrained predictive models, then deployment is simplified, but prediction errors are severe and resource scheduling efficiency deteriorates
Solution Approach 1:
The model undergoes pre-training on a source system before deployment, which simplifies the deployment process by eliminating the need for extensive on-site training. The preliminary training prepares the model in advance, allowing for quick and simple deployment while maintaining resource scheduling efficiency through subsequent parameter adaptation to the target environment.
Solution Approach 2:
Model parameters are adjusted during deployment to adapt the pre-trained model to the target system's specific characteristics. This parameter modification maintains resource scheduling efficiency by ensuring the model accurately predicts resource performance in the target environment, while keeping the deployment process simple and straightforward.
Data Source
AI summary
Systems for distributed resource system management. A first computing system operates in a first computing environment. A predictive model is trained in the first computing environment to form a trained resource performance predictive model that comprises a set of trained model parameters to capture at least computing and storage IO parameters that are responsive to execution of one or more workloads that consume computing and storage resources in the first computing environment. When the trained resource performance predictive model is deployed to a second computing environment, various computing system configuration differences, and/or workload differences and/or other differences between the first computing environment and the second computing environment are detected and measured. Responsive to the detected differences and/or measurements, some of the trained resource performance predictive model parameters are modified to adapt the trained resource performance predictive model to any of the detected and/or measured characteristics of the second computing environment.