AI Workload Management for Asymmetric Link Optimization

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current scalable, universal system interconnects lack a self-learning adaptive engine to dynamically detect relationships between subcomponents and usage metrics, which hinders runtime performance and resiliency in computing ecosystems.

Innovation Solution

Incorporating an AI-fueled analytics engine that monitors resource performance, orchestrates resource allocation, and tunes asymmetric links to optimize lane configurations based on workload demands, using machine learning to create a self-learning, adaptive model for efficient resource management.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If scalable, universal system interconnects are used to unify communications and simplify designs, then interoperability is improved and solution architectures are simplified, but runtime performance and resiliency deteriorate due to lack of adaptive resource management

Engineering Contradiction:

ImproveinteroperabilityVSAvoidruntime performance

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

The patent implements dynamic resource allocation by introducing an AI-fueled analytics engine that continuously monitors workload demands and adjusts network resource allocation in real-time. The system dynamically binds compute and memory resources from fluid pools through fabric switches, transitioning from static to adaptive resource management to resolve the contradiction between universal interoperability and runtime performance

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The patent employs feedback mechanisms through AI-driven analytics that monitor usage metrics and runtime performance, then feed this information back to the resource allocation system. This closed-loop control enables the system to learn from actual performance data and continuously optimize resource distribution, maintaining both interoperability and runtime performance

Inventive Principle:

Principle #23Feedback

2Device complexity

If static network resource allocation is used to simplify management, then device complexity is reduced, but bandwidth utilization deteriorates under varying workload demands

Engineering Contradiction:

Improveresource management complexityVSAvoidbandwidth utilization efficiency

Core Design Contradiction:

Device complexityVSLoss of energy

Solution Approach 1:

The patent implements self-service through autonomous AI-driven resource allocation that automatically monitors workload patterns and adjusts network resource distribution without manual intervention. The system serves itself by learning from usage metrics and making real-time allocation decisions, eliminating the need for complex manual resource management while optimizing bandwidth utilization

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The patent dynamically changes network resource allocation parameters based on workload demands. The AI engine adjusts bandwidth allocation, link configurations, and resource binding parameters in real-time, transforming static resource management into an adaptive system that optimizes bandwidth utilization without proportionally increasing management complexity

Inventive Principle:

Principle #35Parameter changes

3Adaptability or versatility

If fluid pools of compute and memory resources are used to create disaggregated systems, then system composability is improved, but latency increases due to network overhead

Engineering Contradiction:

Improvesystem composabilityVSAvoidruntime latency

Core Design Contradiction:

Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-establishing fabric switch connections and preparing resource binding configurations before workloads are executed. The AI engine anticipates resource allocation needs based on workload patterns, pre-configuring network paths and resource bindings to minimize latency when actual computation occurs

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent implements dynamic resource binding through fabric switches that can rapidly reconfigure network connections based on real-time workload demands. This dynamic binding of compute and memory resources from fluid pools reduces the time required for resource allocation while maintaining system composability and disaggregation

Inventive Principle:

Principle #15Dynamics

Data Source

PatentUS11329890B2Network-aware workload management using artificial intelligence and exploitation of asymmetric link for allocating network resources

Publication Date: 2022.05.10 HEWLETT PACKARD ENTERPRISE DEV LP

AI summary

Systems which support an asymmetric link define rules and policies in each individual physical layer. An asymmetric link is a physical layer with a different number of transmit versus receive lanes. Asymmetric links enable physical layers to optimize performance, power, and system resources based on the required data bandwidth per direction across a link. Modern applications exhibit large demands for high memory bandwidth, i.e., more memory channels and larger bandwidth per channel. The utilization data, patterns) of link usage, and/or patterns) of lane usage may be gathered to exploit the facilities provided by asymmetric links. An engine includes AI-fueled analytics to monitor, orchestrate resources, and provide optimal routing, exploiting asymmetric links, lane polarity, and enqueue-dequeue in a computing ecosystem.