ML-Based Workload Allocation for Accelerator Modules

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Information handling systems face challenges in optimizing workload distribution and resource allocation due to variations in physical and logical arrangements of CXL/PCIe slots, memory riser cards, and DIMMs, which impact performance and efficiency.

Innovation Solution

The system employs a processor with machine learning capabilities to instantiate workloads, determine processing needs, and allocate resources based on the capabilities of accelerator modules, optimizing workload placement and resource allocation through a Compute Express Link (CXL) standard and a BMC that monitors and manages components to achieve maximum processing performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If workload is allocated to accelerator modules based on fixed rules, then system complexity is reduced, but processing performance and efficiency deteriorate due to inability to adapt to variations in physical and logical arrangements

Engineering Contradiction:

Improveworkload allocation complexityVSAvoidprocessing performance

Core Design Contradiction:

Device complexityVSProductivity

Solution Approach 1:

The patent implements dynamic workload allocation by using a machine learning model that continuously learns from system performance data and adapts allocation decisions based on current system state, physical arrangements, and workload characteristics, transforming static rule-based allocation into a dynamic optimization system

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The system changes allocation parameters dynamically by adjusting workload distribution based on learned patterns from historical data, modifying allocation decisions according to variations in physical and logical arrangements of accelerator modules, memory devices, and interconnect topology

Inventive Principle:

Principle #35Parameter changes

2Productivity

If machine learning-based dynamic allocation is implemented, then processing performance and efficiency improve, but system complexity increases due to additional monitoring and decision-making components

Engineering Contradiction:

Improveprocessing performanceVSAvoidsystem architecture complexity

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The system implements self-service through autonomous machine learning models that automatically monitor system state, learn from performance data, and make allocation decisions without human intervention, reducing the need for complex manual configuration and management interfaces

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The patent incorporates feedback mechanisms where performance metrics from workload execution are continuously collected and fed back to the machine learning model, which uses this feedback to refine allocation strategies and improve future decisions, creating a closed-loop optimization system

Inventive Principle:

Principle #23Feedback

3Productivity

If workload placement is optimized for specific physical arrangements, then processing performance improves, but adaptability to different configurations deteriorates

Engineering Contradiction:

Improveprocessing performanceVSAvoidconfiguration adaptability

Core Design Contradiction:

ProductivityVSAdaptability or versatility

Solution Approach 1:

The system performs preliminary learning by training machine learning models on diverse system configurations and workload types before deployment, enabling the model to recognize patterns and make informed allocation decisions across various physical and logical arrangements without requiring reconfiguration

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent creates a universal workload allocation system where the machine learning model serves multiple functions: it adapts to different accelerator module types, handles various memory configurations, and optimizes for diverse workload characteristics, making the system versatile across multiple scenarios

Inventive Principle:

Principle #6Universality (Multi-functionality)

Data Source

PatentUS20240012686A1Workload balance and assignment optimization using machine learining

Publication Date: 2024.01.11 DELL PROD LP

AI summary

An information handling system includes a processor, first and second plug-in connector interfaces coupled to the processor, and first and second accelerator modules installed into respective first and second plug-in connector interfaces. The processor instantiates machine learning code. The information handling system instantiates a workload on the processor. The machine learning code determines a processing need of the workload, determines a first processing capability of the first accelerator module and a second processing capability of the second accelerator module, and allocates a processing resource of the first accelerator module to the workload based upon an evaluation of the processing need, the first processing capability, and the second processing capability.