Biangular Activation Function for Workflow Time Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In shared computing environments, accurately predicting the time-to-finish of workflows is challenging due to the complexity of resource allocation and the need for efficient scalability, particularly for Infrastructure and Platform as a Service providers, where existing methods lack precision and are costly.
Innovation Solution
Employing a deep neural network with a biangular activation function, such as the Biangular Leaky Rectified Linear Unit (BILU), to predict the time-to-finish of workflows by learning from past executions and adapting to new instances, using a combination of provenance data and telemetry information to optimize resource allocation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a deep neural network with conventional activation functions is used to predict workflow time-to-finish, then the model can capture complex patterns from past executions, but the number of parameters increases leading to higher computational complexity and slower training
Solution Approach 1:
The patent changes the parameters of the activation function by introducing a biangular activation function with only four learnable parameters (two angles and their corresponding slopes) instead of the numerous parameters in conventional activation functions. This parameter reduction directly addresses the contradiction by maintaining prediction accuracy while significantly reducing model complexity and training time.
Solution Approach 2:
The biangular activation function serves multiple functions: it acts as a universal approximation function that can model complex non-linear relationships in workflow execution data, while simultaneously serving as a parameter-efficient alternative to complex activation functions. This multi-functionality resolves the contradiction by achieving both high prediction accuracy and low model complexity.
2Measurement precision
If more parameters are used in the activation function to improve prediction accuracy, then the model can better fit the workflow execution patterns, but the training time and computational resources increase
Solution Approach 1:
The patent reduces the number of parameters in the activation function from potentially hundreds or thousands in conventional functions to just four parameters (two angles and their slopes). This dramatic parameter reduction directly decreases training time and computational resource requirements while maintaining the ability to achieve high prediction accuracy through efficient learning of these fewer parameters.
3Measurement precision
If a complex model is used to accurately predict time-to-finish for different workflow configurations, then prediction precision improves, but the model becomes less adaptable to new workflow instances
Solution Approach 1:
The biangular activation function with its four parameters provides a balance between model capacity and adaptability. The parameters can be efficiently learned during training and quickly adapted to new workflow instances, whereas complex models with many parameters require extensive training data and time to achieve similar adaptability. This parameter-efficient design directly addresses the contradiction.
Solution Approach 2:
The activation function parameters (angles and slopes) are designed to be dynamically learnable and adaptable to different workflow characteristics. This dynamic parameter adjustment capability allows the model to efficiently adapt to new workflow instances without requiring a complete retraining process, thus resolving the contradiction between accuracy and adaptability.
Data Source
AI summary
Techniques are provided for predicting a time-to-finish of at least one workflow in a shared computing environment using a deep neural network with a biangular activation function. An exemplary method comprises: obtaining a specification of an executing workflow of multiple concurrent workflows in a shared computing environment, wherein the specification comprises states of past executions of the executing workflow; obtaining a trained deep neural network, wherein the trained deep neural network is trained to predict one or more future states of the executing workflow using the states of past executions and wherein the trained deep neural network employs a biangular activation function comprising multiple parameters that define a position and a slope associated with two angles of the biangular activation function for a range of input values; and estimating, using the at least one trained deep neural network, a time-to-finish of the executing workflow of the multiple concurrent workflows.