Virtual GPU Capacity Reservation via Segmented Availability

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

In large-scale distributed systems, managing concurrent requests for virtual graphics processing units (GPUs) leads to contention and inefficiency due to the lack of effective resource reservation mechanisms, particularly in multi-tenant environments where multiple clients require varying graphics processing capabilities and locations.

Innovation Solution

Implementing a capacity reservation system that segments data store availability information by constraints such as virtual GPU type and location, allowing for random selection within segments and pages to reduce contention, ensuring successful reservation of virtual GPUs for virtual compute instances.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a centralized resource reservation system is used for virtual GPUs, then resource allocation control is improved, but contention and inefficiency increase due to concurrent requests from multiple clients

Engineering Contradiction:
Improveresource allocation controlVSAvoidconcurrent request handling efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent segments the virtual GPU resource pool into multiple discrete resource units, each with specific configurations (GPU type, memory, compute capacity). This segmentation allows the reservation system to allocate specific units to different clients rather than managing a monolithic resource pool, reducing contention by enabling parallel allocation decisions for multiple clients simultaneously.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements preliminary capacity verification and reservation mechanisms where clients can check availability and reserve virtual GPU resources before actually needing them. The system performs capacity verification in advance, maintains reservation states, and allows clients to secure resources ahead of time, eliminating the need for all concurrent requests to compete for immediately available resources.

Inventive Principle:
Principle #10Preliminary action

2Productivity

If virtualization technologies are used to share physical computing devices among multiple users, then resource utilization is improved, but managing concurrent requests leads to contention and inefficiency

Engineering Contradiction:
Improveresource utilizationVSAvoidconcurrent request management
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent applies local quality by allowing different clients to access virtual GPU resources with different local characteristics and constraints. Each client's reservation request can specify local requirements (GPU type, location, capacity), and the system satisfies each request with resources matching those local qualities rather than providing uniform access to all resources.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent introduces an intermediary reservation system that sits between the physical GPU resources and multiple virtualized clients. This intermediary layer manages capacity verification, maintains availability information, and coordinates allocations, simplifying the management of concurrent requests by providing a centralized coordination point rather than allowing direct client-to-resource contention.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Measurement precision

If capacity verification is performed for each concurrent request, then resource allocation accuracy is improved, but processing time and contention increase

Engineering Contradiction:
Improvecapacity verification accuracyVSAvoidrequest processing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent merges multiple capacity verification operations by maintaining a centralized availability information structure that tracks the state of all virtual GPU resources. Instead of performing independent verification for each request, the system updates and queries this merged state, allowing multiple requests to be processed against a unified view of resource availability, reducing redundant verification work.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent performs preliminary capacity verification by maintaining up-to-date availability information in advance of actual allocation requests. The system proactively tracks resource states, updates availability as resources are allocated or released, and makes this information readily accessible, so that when concurrent requests arrive, the verification step is minimized to a simple query rather than a full capacity analysis.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3552103B1Capacity reservation for virtualized graphics processing
Publication Date: 2023.09.06 AMAZON TECH INC
  • EP3552103B1 patent drawingFigure 1
  • EP3552103B1 patent drawingFigure 2A
  • EP3552103B1 patent drawingFigure 2B

AI summary

Methods, systems, and computer-readable media for capacity reservation for virtualized graphics processing are disclosed. A request is received to attach a virtual GPU to a virtual compute instance. The request comprises one or more constraints. Availability information is retrieved from a data store that indicates virtual GPUs available in a provider network and matching the one or more constraints. A virtual GPU is selected from among the available virtual GPUs in the availability information. The selected virtual GPU is reserved for attachment to the virtual compute instance. The virtual compute instance is implemented using CPU resources and memory resources of a physical compute instance, the virtual GPU is implemented using a physical GPU in the provider network, and the physical GPU is accessible to the physical compute instance over a network.