Processing Resource Allocation Using Predictive Saturation Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In parallel processing systems, allocating processing resources to nodes that exhibit diminishing returns is challenging, as increasing resources lead to smaller improvements in processing outcomes, and existing methods like Amdahl's law do not optimize resource allocation between multiple nodes for maximizing total processing outcomes.
Innovation Solution
Training predictive models for each processing node, implemented as monotonically increasing functions, to identify optimal resource allocation that maximizes total processing outcomes, using concave piecewise linear, isotonic, or sigmoid models, and solving optimization problems using linear programming or mixed integer linear programming techniques.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If processing resources are increased to improve processing outcomes, then processing outcomes are improved, but the improvement diminishes with each additional resource allocated
Solution Approach 1:
The patent applies parameter changes by transforming the resource allocation problem into an optimization problem where the parameters (resource allocation amounts) are adjusted to maximize processing outcomes. The system changes the allocation parameters dynamically based on predictive models that capture the diminishing returns characteristic, identifying optimal allocation points where marginal gains are maximized rather than continuously increasing resources.
2Productivity
If processing resources are allocated to multiple processing nodes, then total processing outcomes can be maximized, but determining optimal allocation becomes computationally complex
Solution Approach 1:
The patent applies segmentation by dividing the overall resource allocation problem into individual predictive models for each processing node. Each node's resource-outcome relationship is modeled separately, allowing the complex multi-node optimization to be decomposed into manageable components that can be solved independently and then combined to achieve the global optimum.
3Productivity
If existing methods like Amdahl's law are used for resource allocation, then some optimization is achieved, but they do not optimize resource allocation between multiple nodes for maximizing total processing outcomes
Solution Approach 1:
The patent applies universality by creating a generalizable framework using predictive models that can be applied to any number of processing nodes with different diminishing returns characteristics. The optimization approach is versatile and adapts to various node configurations and resource types, making it universally applicable beyond single-node scenarios while maximizing total processing outcomes across the system.
Data Source
AI summary
There is provided a computer implemented method of allocating processing resources for processing by processing nodes, comprising: training predictive models, each predictive model for a respective processing node, each predictive model trained on a training dataset comprising records, each record including a historical amount of processing resources allocated to the respective processing node and a ground truth label indicating historical processing outcomes, wherein each processing node exhibits diminishing returns of processing outcomes with increasing allocated processing resources, wherein each predictive model is implemented as a monotonically increasing function that reaches a saturation level, solving an optimization allocation problem using the predictive models to identify a respective amount of processing resources for allocation to each processing node that maximizes a total of processing outcomes for a predetermined total amount of processing resources, and generating instructions for allocation of the respective amount of processing resources to each respective processing node.