Transferable Clustering of Contextual Bandits for Cloud Resource Allocation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Resource allocation for cloud services, particularly for GPU instances, is challenging due to the high computational costs and limited availability of offline data, leading to over-provisioning and resource wastage.
Innovation Solution
A transfer learning-based approach is applied, using historical CPU instance data to learn optimal GPU usage by formulating the cross-domain resource allocation problem as an online contextual bandit problem and employing a transferable clustering of bandits algorithm.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If extensive experimentation is performed to determine optimal GPU resource allocation, then resource allocation accuracy is improved, but computational cost and time consumption increase significantly
Solution Approach 1:
The patent performs preliminary action by pre-training a source domain model on abundant CPU instance data before deploying to the target GPU domain. This pre-training phase captures general resource allocation patterns and service characteristics, which are then transferred to the GPU domain through domain adaptation, eliminating the need for extensive experimentation on expensive GPU instances.
Solution Approach 2:
The patent applies copying by creating a source domain model that replicates the resource allocation optimization capabilities for CPU instances, then transferring this model to the GPU domain. The source domain model serves as a copy that can be adapted to the target domain, allowing optimal GPU resource allocation to be determined without direct experimentation.
2Productivity
If domain adaptation is performed to transfer CPU data to GPU domain, then resource allocation efficiency is improved, but model complexity increases
Solution Approach 1:
The patent applies parameter changes by modifying the source domain model's parameters through domain adaptation techniques. The model structure remains the same, but the parameters are adjusted and retrained on transformed GPU domain data, allowing efficient transfer of knowledge from CPU to GPU domain while maintaining model manageability.
Solution Approach 2:
The patent uses an intermediary approach by introducing a domain adapter layer between the source domain model and the target domain. This intermediary component facilitates the transfer of knowledge from CPU instances to GPU instances, enabling domain adaptation without requiring complete model redesign and reducing overall system complexity.
3Quantity of substance
If transfer learning is used to leverage CPU data for GPU allocation, then data utilization efficiency is improved, but measurement and detection difficulty increases
Solution Approach 1:
The patent replaces the mechanical system of direct GPU experimentation with a computational transfer learning system. Instead of physically experimenting with GPU resources, the system uses algorithms to transform and adapt CPU domain data into the GPU domain, substituting physical experimentation with computational transformation.
Solution Approach 2:
The patent applies parameter changes by transforming the feature parameters and data distribution from the CPU domain to the GPU domain through domain adaptation techniques. This transformation allows the model to leverage abundant CPU data while adapting it to the target GPU domain, improving data utilization efficiency.
Data Source
AI summary
Methods for determining optimal cloud service resource include determining a reward function for a set of resource configurations identifying cloud service resource parameters. The cloud service resource parameters include a source parameter and a target parameter of services to provide a client computing device. A source parameter dataset for the source parameter and a target parameter dataset is generated using the reward function and historical source parameter data. The matrices are then subject to SVD and clustering. A target parameter reward dataset is learned from output of the SVD and clustering. The target parameter dataset is used to determine the parameters for the target parameter for providing corresponding cloud service resources.