Memory-Aware vGPU Placement via ILP Optimization

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing virtual GPU (vGPU) enabled systems face sub-optimal virtual machine placement due to insufficient consideration of specific GPU resource demands, leading to inefficient resource utilization and network saturation.

Innovation Solution

The implementation of memory-aware vGPU request placement models, such as Integer Linear Programming (ILP) and first fit heuristics, that take into account GPU memory requirements and vGPU profiles to optimize the assignment of vGPU requests to GPUs, minimizing the number of GPUs utilized and ensuring sufficient memory reservations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If existing placement solutions are used that do not consider GPU memory requirements, then the placement process is simple, but resource utilization is inefficient and network saturation occurs

Engineering Contradiction:

Improveresource utilization efficiencyVSAvoidplacement model complexity

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The patent transforms the placement problem from a simple resource matching task to an optimized resource allocation problem by introducing GPU memory requirements as a critical parameter. The ILP formulation incorporates memory constraints, vGPU profiles, and placement decisions as variables, fundamentally changing how placement is performed to achieve optimal resource utilization while preventing network saturation

Inventive Principle:

Principle #35Parameter changes

Solution Approach 2:

The patent performs preliminary analysis of GPU memory requirements and vGPU profiles before making placement decisions. By pre-processing and understanding the memory characteristics of virtual machines and available GPU resources, the system can make informed placement decisions that optimize resource utilization from the outset, rather than reacting to overload conditions

Inventive Principle:

Principle #10Preliminary action

2Manufacturing precision

If more information about GPU resource demands is collected and used, then placement optimization improves, but the complexity of the placement decision process increases

Engineering Contradiction:

Improveplacement optimization precisionVSAvoidplacement decision complexity

Core Design Contradiction:

Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent segments the placement decision into distinct components: virtual machine requirements analysis, GPU resource availability assessment, constraint formulation, and optimization execution. By dividing the complex decision process into manageable segments, the system can incorporate detailed GPU memory information without overwhelming complexity, as each segment handles a specific aspect of the placement problem

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent introduces an intermediary optimization layer (ILP formulation) that acts as a mediator between raw placement data and final placement decisions. This intermediary layer processes detailed GPU memory requirements and vGPU profile information, transforming them into optimized placement decisions while managing the complexity of the decision-making process

Inventive Principle:

Principle #24Intermediary (Mediator)

3Reliability

If virtual machines are relocated to balance resource utilization, then host overload is reduced, but network saturation and link overload may occur

Engineering Contradiction:

Improvehost resource balanceVSAvoidnetwork saturation

Core Design Contradiction:

ReliabilityVSObject-generated harmful factors

Solution Approach 1:

The patent performs preliminary placement optimization before host overload occurs by considering GPU memory requirements and network topology in the placement decision. By proactively placing virtual machines on appropriate hosts based on detailed resource matching, the system prevents the need for reactive VM relocation that would cause network saturation

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent incorporates feedback mechanisms that monitor resource utilization and placement effectiveness. By continuously assessing whether placements are meeting GPU memory requirements and resource balance goals, the system can adjust future placement decisions to maintain host balance without triggering network saturation from frequent VM migrations

Inventive Principle:

Principle #23Feedback

Data Source

PatentUS11263054B2Memory-aware placement for virtual GPU enabled systems

Publication Date: 2022.03.01 VMWARE INC

AI summary

Disclosed are aspects of memory-aware placement in systems that include graphics processing units (GPUs) that are virtual GPU (vGPU) enabled. In some embodiments, a computing environment is monitored to identify graphics processing unit (GPU) data for a plurality of virtual GPU (vGPU) enabled GPUs of the computing environment, a plurality of vGPU requests are received. A respective vGPU request includes a GPU memory requirement. GPU configurations are determined in order to accommodate vGPU requests. The GPU configurations are determined based on an integer linear programming (ILP) vGPU request placement model. Configured vGPU profiles are applied for vGPU enabled GPUs, and vGPUs are created based on the configured vGPU profiles. The vGPU requests are assigned to the vGPUs.